Loading...
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#include <linux/sched/signal.h>
37
38#define MLOG_MASK_PREFIX ML_DLM_GLUE
39#include <cluster/masklog.h>
40
41#include "ocfs2.h"
42#include "ocfs2_lockingver.h"
43
44#include "alloc.h"
45#include "dcache.h"
46#include "dlmglue.h"
47#include "extent_map.h"
48#include "file.h"
49#include "heartbeat.h"
50#include "inode.h"
51#include "journal.h"
52#include "stackglue.h"
53#include "slot_map.h"
54#include "super.h"
55#include "uptodate.h"
56#include "quota.h"
57#include "refcounttree.h"
58#include "acl.h"
59
60#include "buffer_head_io.h"
61
62struct ocfs2_mask_waiter {
63 struct list_head mw_item;
64 int mw_status;
65 struct completion mw_complete;
66 unsigned long mw_mask;
67 unsigned long mw_goal;
68#ifdef CONFIG_OCFS2_FS_STATS
69 ktime_t mw_lock_start;
70#endif
71};
72
73static struct ocfs2_super *ocfs2_get_dentry_osb(struct ocfs2_lock_res *lockres);
74static struct ocfs2_super *ocfs2_get_inode_osb(struct ocfs2_lock_res *lockres);
75static struct ocfs2_super *ocfs2_get_file_osb(struct ocfs2_lock_res *lockres);
76static struct ocfs2_super *ocfs2_get_qinfo_osb(struct ocfs2_lock_res *lockres);
77
78/*
79 * Return value from ->downconvert_worker functions.
80 *
81 * These control the precise actions of ocfs2_unblock_lock()
82 * and ocfs2_process_blocked_lock()
83 *
84 */
85enum ocfs2_unblock_action {
86 UNBLOCK_CONTINUE = 0, /* Continue downconvert */
87 UNBLOCK_CONTINUE_POST = 1, /* Continue downconvert, fire
88 * ->post_unlock callback */
89 UNBLOCK_STOP_POST = 2, /* Do not downconvert, fire
90 * ->post_unlock() callback. */
91};
92
93struct ocfs2_unblock_ctl {
94 int requeue;
95 enum ocfs2_unblock_action unblock_action;
96};
97
98/* Lockdep class keys */
99struct lock_class_key lockdep_keys[OCFS2_NUM_LOCK_TYPES];
100
101static int ocfs2_check_meta_downconvert(struct ocfs2_lock_res *lockres,
102 int new_level);
103static void ocfs2_set_meta_lvb(struct ocfs2_lock_res *lockres);
104
105static int ocfs2_data_convert_worker(struct ocfs2_lock_res *lockres,
106 int blocking);
107
108static int ocfs2_dentry_convert_worker(struct ocfs2_lock_res *lockres,
109 int blocking);
110
111static void ocfs2_dentry_post_unlock(struct ocfs2_super *osb,
112 struct ocfs2_lock_res *lockres);
113
114static void ocfs2_set_qinfo_lvb(struct ocfs2_lock_res *lockres);
115
116static int ocfs2_check_refcount_downconvert(struct ocfs2_lock_res *lockres,
117 int new_level);
118static int ocfs2_refcount_convert_worker(struct ocfs2_lock_res *lockres,
119 int blocking);
120
121#define mlog_meta_lvb(__level, __lockres) ocfs2_dump_meta_lvb_info(__level, __PRETTY_FUNCTION__, __LINE__, __lockres)
122
123/* This aids in debugging situations where a bad LVB might be involved. */
124static void ocfs2_dump_meta_lvb_info(u64 level,
125 const char *function,
126 unsigned int line,
127 struct ocfs2_lock_res *lockres)
128{
129 struct ocfs2_meta_lvb *lvb = ocfs2_dlm_lvb(&lockres->l_lksb);
130
131 mlog(level, "LVB information for %s (called from %s:%u):\n",
132 lockres->l_name, function, line);
133 mlog(level, "version: %u, clusters: %u, generation: 0x%x\n",
134 lvb->lvb_version, be32_to_cpu(lvb->lvb_iclusters),
135 be32_to_cpu(lvb->lvb_igeneration));
136 mlog(level, "size: %llu, uid %u, gid %u, mode 0x%x\n",
137 (unsigned long long)be64_to_cpu(lvb->lvb_isize),
138 be32_to_cpu(lvb->lvb_iuid), be32_to_cpu(lvb->lvb_igid),
139 be16_to_cpu(lvb->lvb_imode));
140 mlog(level, "nlink %u, atime_packed 0x%llx, ctime_packed 0x%llx, "
141 "mtime_packed 0x%llx iattr 0x%x\n", be16_to_cpu(lvb->lvb_inlink),
142 (long long)be64_to_cpu(lvb->lvb_iatime_packed),
143 (long long)be64_to_cpu(lvb->lvb_ictime_packed),
144 (long long)be64_to_cpu(lvb->lvb_imtime_packed),
145 be32_to_cpu(lvb->lvb_iattr));
146}
147
148
149/*
150 * OCFS2 Lock Resource Operations
151 *
152 * These fine tune the behavior of the generic dlmglue locking infrastructure.
153 *
154 * The most basic of lock types can point ->l_priv to their respective
155 * struct ocfs2_super and allow the default actions to manage things.
156 *
157 * Right now, each lock type also needs to implement an init function,
158 * and trivial lock/unlock wrappers. ocfs2_simple_drop_lockres()
159 * should be called when the lock is no longer needed (i.e., object
160 * destruction time).
161 */
162struct ocfs2_lock_res_ops {
163 /*
164 * Translate an ocfs2_lock_res * into an ocfs2_super *. Define
165 * this callback if ->l_priv is not an ocfs2_super pointer
166 */
167 struct ocfs2_super * (*get_osb)(struct ocfs2_lock_res *);
168
169 /*
170 * Optionally called in the downconvert thread after a
171 * successful downconvert. The lockres will not be referenced
172 * after this callback is called, so it is safe to free
173 * memory, etc.
174 *
175 * The exact semantics of when this is called are controlled
176 * by ->downconvert_worker()
177 */
178 void (*post_unlock)(struct ocfs2_super *, struct ocfs2_lock_res *);
179
180 /*
181 * Allow a lock type to add checks to determine whether it is
182 * safe to downconvert a lock. Return 0 to re-queue the
183 * downconvert at a later time, nonzero to continue.
184 *
185 * For most locks, the default checks that there are no
186 * incompatible holders are sufficient.
187 *
188 * Called with the lockres spinlock held.
189 */
190 int (*check_downconvert)(struct ocfs2_lock_res *, int);
191
192 /*
193 * Allows a lock type to populate the lock value block. This
194 * is called on downconvert, and when we drop a lock.
195 *
196 * Locks that want to use this should set LOCK_TYPE_USES_LVB
197 * in the flags field.
198 *
199 * Called with the lockres spinlock held.
200 */
201 void (*set_lvb)(struct ocfs2_lock_res *);
202
203 /*
204 * Called from the downconvert thread when it is determined
205 * that a lock will be downconverted. This is called without
206 * any locks held so the function can do work that might
207 * schedule (syncing out data, etc).
208 *
209 * This should return any one of the ocfs2_unblock_action
210 * values, depending on what it wants the thread to do.
211 */
212 int (*downconvert_worker)(struct ocfs2_lock_res *, int);
213
214 /*
215 * LOCK_TYPE_* flags which describe the specific requirements
216 * of a lock type. Descriptions of each individual flag follow.
217 */
218 int flags;
219};
220
221/*
222 * Some locks want to "refresh" potentially stale data when a
223 * meaningful (PRMODE or EXMODE) lock level is first obtained. If this
224 * flag is set, the OCFS2_LOCK_NEEDS_REFRESH flag will be set on the
225 * individual lockres l_flags member from the ast function. It is
226 * expected that the locking wrapper will clear the
227 * OCFS2_LOCK_NEEDS_REFRESH flag when done.
228 */
229#define LOCK_TYPE_REQUIRES_REFRESH 0x1
230
231/*
232 * Indicate that a lock type makes use of the lock value block. The
233 * ->set_lvb lock type callback must be defined.
234 */
235#define LOCK_TYPE_USES_LVB 0x2
236
237static struct ocfs2_lock_res_ops ocfs2_inode_rw_lops = {
238 .get_osb = ocfs2_get_inode_osb,
239 .flags = 0,
240};
241
242static struct ocfs2_lock_res_ops ocfs2_inode_inode_lops = {
243 .get_osb = ocfs2_get_inode_osb,
244 .check_downconvert = ocfs2_check_meta_downconvert,
245 .set_lvb = ocfs2_set_meta_lvb,
246 .downconvert_worker = ocfs2_data_convert_worker,
247 .flags = LOCK_TYPE_REQUIRES_REFRESH|LOCK_TYPE_USES_LVB,
248};
249
250static struct ocfs2_lock_res_ops ocfs2_super_lops = {
251 .flags = LOCK_TYPE_REQUIRES_REFRESH,
252};
253
254static struct ocfs2_lock_res_ops ocfs2_rename_lops = {
255 .flags = 0,
256};
257
258static struct ocfs2_lock_res_ops ocfs2_nfs_sync_lops = {
259 .flags = 0,
260};
261
262static struct ocfs2_lock_res_ops ocfs2_trim_fs_lops = {
263 .flags = LOCK_TYPE_REQUIRES_REFRESH|LOCK_TYPE_USES_LVB,
264};
265
266static struct ocfs2_lock_res_ops ocfs2_orphan_scan_lops = {
267 .flags = LOCK_TYPE_REQUIRES_REFRESH|LOCK_TYPE_USES_LVB,
268};
269
270static struct ocfs2_lock_res_ops ocfs2_dentry_lops = {
271 .get_osb = ocfs2_get_dentry_osb,
272 .post_unlock = ocfs2_dentry_post_unlock,
273 .downconvert_worker = ocfs2_dentry_convert_worker,
274 .flags = 0,
275};
276
277static struct ocfs2_lock_res_ops ocfs2_inode_open_lops = {
278 .get_osb = ocfs2_get_inode_osb,
279 .flags = 0,
280};
281
282static struct ocfs2_lock_res_ops ocfs2_flock_lops = {
283 .get_osb = ocfs2_get_file_osb,
284 .flags = 0,
285};
286
287static struct ocfs2_lock_res_ops ocfs2_qinfo_lops = {
288 .set_lvb = ocfs2_set_qinfo_lvb,
289 .get_osb = ocfs2_get_qinfo_osb,
290 .flags = LOCK_TYPE_REQUIRES_REFRESH | LOCK_TYPE_USES_LVB,
291};
292
293static struct ocfs2_lock_res_ops ocfs2_refcount_block_lops = {
294 .check_downconvert = ocfs2_check_refcount_downconvert,
295 .downconvert_worker = ocfs2_refcount_convert_worker,
296 .flags = 0,
297};
298
299static inline int ocfs2_is_inode_lock(struct ocfs2_lock_res *lockres)
300{
301 return lockres->l_type == OCFS2_LOCK_TYPE_META ||
302 lockres->l_type == OCFS2_LOCK_TYPE_RW ||
303 lockres->l_type == OCFS2_LOCK_TYPE_OPEN;
304}
305
306static inline struct ocfs2_lock_res *ocfs2_lksb_to_lock_res(struct ocfs2_dlm_lksb *lksb)
307{
308 return container_of(lksb, struct ocfs2_lock_res, l_lksb);
309}
310
311static inline struct inode *ocfs2_lock_res_inode(struct ocfs2_lock_res *lockres)
312{
313 BUG_ON(!ocfs2_is_inode_lock(lockres));
314
315 return (struct inode *) lockres->l_priv;
316}
317
318static inline struct ocfs2_dentry_lock *ocfs2_lock_res_dl(struct ocfs2_lock_res *lockres)
319{
320 BUG_ON(lockres->l_type != OCFS2_LOCK_TYPE_DENTRY);
321
322 return (struct ocfs2_dentry_lock *)lockres->l_priv;
323}
324
325static inline struct ocfs2_mem_dqinfo *ocfs2_lock_res_qinfo(struct ocfs2_lock_res *lockres)
326{
327 BUG_ON(lockres->l_type != OCFS2_LOCK_TYPE_QINFO);
328
329 return (struct ocfs2_mem_dqinfo *)lockres->l_priv;
330}
331
332static inline struct ocfs2_refcount_tree *
333ocfs2_lock_res_refcount_tree(struct ocfs2_lock_res *res)
334{
335 return container_of(res, struct ocfs2_refcount_tree, rf_lockres);
336}
337
338static inline struct ocfs2_super *ocfs2_get_lockres_osb(struct ocfs2_lock_res *lockres)
339{
340 if (lockres->l_ops->get_osb)
341 return lockres->l_ops->get_osb(lockres);
342
343 return (struct ocfs2_super *)lockres->l_priv;
344}
345
346static int ocfs2_lock_create(struct ocfs2_super *osb,
347 struct ocfs2_lock_res *lockres,
348 int level,
349 u32 dlm_flags);
350static inline int ocfs2_may_continue_on_blocked_lock(struct ocfs2_lock_res *lockres,
351 int wanted);
352static void __ocfs2_cluster_unlock(struct ocfs2_super *osb,
353 struct ocfs2_lock_res *lockres,
354 int level, unsigned long caller_ip);
355static inline void ocfs2_cluster_unlock(struct ocfs2_super *osb,
356 struct ocfs2_lock_res *lockres,
357 int level)
358{
359 __ocfs2_cluster_unlock(osb, lockres, level, _RET_IP_);
360}
361
362static inline void ocfs2_generic_handle_downconvert_action(struct ocfs2_lock_res *lockres);
363static inline void ocfs2_generic_handle_convert_action(struct ocfs2_lock_res *lockres);
364static inline void ocfs2_generic_handle_attach_action(struct ocfs2_lock_res *lockres);
365static int ocfs2_generic_handle_bast(struct ocfs2_lock_res *lockres, int level);
366static void ocfs2_schedule_blocked_lock(struct ocfs2_super *osb,
367 struct ocfs2_lock_res *lockres);
368static inline void ocfs2_recover_from_dlm_error(struct ocfs2_lock_res *lockres,
369 int convert);
370#define ocfs2_log_dlm_error(_func, _err, _lockres) do { \
371 if ((_lockres)->l_type != OCFS2_LOCK_TYPE_DENTRY) \
372 mlog(ML_ERROR, "DLM error %d while calling %s on resource %s\n", \
373 _err, _func, _lockres->l_name); \
374 else \
375 mlog(ML_ERROR, "DLM error %d while calling %s on resource %.*s%08x\n", \
376 _err, _func, OCFS2_DENTRY_LOCK_INO_START - 1, (_lockres)->l_name, \
377 (unsigned int)ocfs2_get_dentry_lock_ino(_lockres)); \
378} while (0)
379static int ocfs2_downconvert_thread(void *arg);
380static void ocfs2_downconvert_on_unlock(struct ocfs2_super *osb,
381 struct ocfs2_lock_res *lockres);
382static int ocfs2_inode_lock_update(struct inode *inode,
383 struct buffer_head **bh);
384static void ocfs2_drop_osb_locks(struct ocfs2_super *osb);
385static inline int ocfs2_highest_compat_lock_level(int level);
386static unsigned int ocfs2_prepare_downconvert(struct ocfs2_lock_res *lockres,
387 int new_level);
388static int ocfs2_downconvert_lock(struct ocfs2_super *osb,
389 struct ocfs2_lock_res *lockres,
390 int new_level,
391 int lvb,
392 unsigned int generation);
393static int ocfs2_prepare_cancel_convert(struct ocfs2_super *osb,
394 struct ocfs2_lock_res *lockres);
395static int ocfs2_cancel_convert(struct ocfs2_super *osb,
396 struct ocfs2_lock_res *lockres);
397
398
399static void ocfs2_build_lock_name(enum ocfs2_lock_type type,
400 u64 blkno,
401 u32 generation,
402 char *name)
403{
404 int len;
405
406 BUG_ON(type >= OCFS2_NUM_LOCK_TYPES);
407
408 len = snprintf(name, OCFS2_LOCK_ID_MAX_LEN, "%c%s%016llx%08x",
409 ocfs2_lock_type_char(type), OCFS2_LOCK_ID_PAD,
410 (long long)blkno, generation);
411
412 BUG_ON(len != (OCFS2_LOCK_ID_MAX_LEN - 1));
413
414 mlog(0, "built lock resource with name: %s\n", name);
415}
416
417static DEFINE_SPINLOCK(ocfs2_dlm_tracking_lock);
418
419static void ocfs2_add_lockres_tracking(struct ocfs2_lock_res *res,
420 struct ocfs2_dlm_debug *dlm_debug)
421{
422 mlog(0, "Add tracking for lockres %s\n", res->l_name);
423
424 spin_lock(&ocfs2_dlm_tracking_lock);
425 list_add(&res->l_debug_list, &dlm_debug->d_lockres_tracking);
426 spin_unlock(&ocfs2_dlm_tracking_lock);
427}
428
429static void ocfs2_remove_lockres_tracking(struct ocfs2_lock_res *res)
430{
431 spin_lock(&ocfs2_dlm_tracking_lock);
432 if (!list_empty(&res->l_debug_list))
433 list_del_init(&res->l_debug_list);
434 spin_unlock(&ocfs2_dlm_tracking_lock);
435}
436
437#ifdef CONFIG_OCFS2_FS_STATS
438static void ocfs2_init_lock_stats(struct ocfs2_lock_res *res)
439{
440 res->l_lock_refresh = 0;
441 memset(&res->l_lock_prmode, 0, sizeof(struct ocfs2_lock_stats));
442 memset(&res->l_lock_exmode, 0, sizeof(struct ocfs2_lock_stats));
443}
444
445static void ocfs2_update_lock_stats(struct ocfs2_lock_res *res, int level,
446 struct ocfs2_mask_waiter *mw, int ret)
447{
448 u32 usec;
449 ktime_t kt;
450 struct ocfs2_lock_stats *stats;
451
452 if (level == LKM_PRMODE)
453 stats = &res->l_lock_prmode;
454 else if (level == LKM_EXMODE)
455 stats = &res->l_lock_exmode;
456 else
457 return;
458
459 kt = ktime_sub(ktime_get(), mw->mw_lock_start);
460 usec = ktime_to_us(kt);
461
462 stats->ls_gets++;
463 stats->ls_total += ktime_to_ns(kt);
464 /* overflow */
465 if (unlikely(stats->ls_gets == 0)) {
466 stats->ls_gets++;
467 stats->ls_total = ktime_to_ns(kt);
468 }
469
470 if (stats->ls_max < usec)
471 stats->ls_max = usec;
472
473 if (ret)
474 stats->ls_fail++;
475}
476
477static inline void ocfs2_track_lock_refresh(struct ocfs2_lock_res *lockres)
478{
479 lockres->l_lock_refresh++;
480}
481
482static inline void ocfs2_init_start_time(struct ocfs2_mask_waiter *mw)
483{
484 mw->mw_lock_start = ktime_get();
485}
486#else
487static inline void ocfs2_init_lock_stats(struct ocfs2_lock_res *res)
488{
489}
490static inline void ocfs2_update_lock_stats(struct ocfs2_lock_res *res,
491 int level, struct ocfs2_mask_waiter *mw, int ret)
492{
493}
494static inline void ocfs2_track_lock_refresh(struct ocfs2_lock_res *lockres)
495{
496}
497static inline void ocfs2_init_start_time(struct ocfs2_mask_waiter *mw)
498{
499}
500#endif
501
502static void ocfs2_lock_res_init_common(struct ocfs2_super *osb,
503 struct ocfs2_lock_res *res,
504 enum ocfs2_lock_type type,
505 struct ocfs2_lock_res_ops *ops,
506 void *priv)
507{
508 res->l_type = type;
509 res->l_ops = ops;
510 res->l_priv = priv;
511
512 res->l_level = DLM_LOCK_IV;
513 res->l_requested = DLM_LOCK_IV;
514 res->l_blocking = DLM_LOCK_IV;
515 res->l_action = OCFS2_AST_INVALID;
516 res->l_unlock_action = OCFS2_UNLOCK_INVALID;
517
518 res->l_flags = OCFS2_LOCK_INITIALIZED;
519
520 ocfs2_add_lockres_tracking(res, osb->osb_dlm_debug);
521
522 ocfs2_init_lock_stats(res);
523#ifdef CONFIG_DEBUG_LOCK_ALLOC
524 if (type != OCFS2_LOCK_TYPE_OPEN)
525 lockdep_init_map(&res->l_lockdep_map, ocfs2_lock_type_strings[type],
526 &lockdep_keys[type], 0);
527 else
528 res->l_lockdep_map.key = NULL;
529#endif
530}
531
532void ocfs2_lock_res_init_once(struct ocfs2_lock_res *res)
533{
534 /* This also clears out the lock status block */
535 memset(res, 0, sizeof(struct ocfs2_lock_res));
536 spin_lock_init(&res->l_lock);
537 init_waitqueue_head(&res->l_event);
538 INIT_LIST_HEAD(&res->l_blocked_list);
539 INIT_LIST_HEAD(&res->l_mask_waiters);
540 INIT_LIST_HEAD(&res->l_holders);
541}
542
543void ocfs2_inode_lock_res_init(struct ocfs2_lock_res *res,
544 enum ocfs2_lock_type type,
545 unsigned int generation,
546 struct inode *inode)
547{
548 struct ocfs2_lock_res_ops *ops;
549
550 switch(type) {
551 case OCFS2_LOCK_TYPE_RW:
552 ops = &ocfs2_inode_rw_lops;
553 break;
554 case OCFS2_LOCK_TYPE_META:
555 ops = &ocfs2_inode_inode_lops;
556 break;
557 case OCFS2_LOCK_TYPE_OPEN:
558 ops = &ocfs2_inode_open_lops;
559 break;
560 default:
561 mlog_bug_on_msg(1, "type: %d\n", type);
562 ops = NULL; /* thanks, gcc */
563 break;
564 };
565
566 ocfs2_build_lock_name(type, OCFS2_I(inode)->ip_blkno,
567 generation, res->l_name);
568 ocfs2_lock_res_init_common(OCFS2_SB(inode->i_sb), res, type, ops, inode);
569}
570
571static struct ocfs2_super *ocfs2_get_inode_osb(struct ocfs2_lock_res *lockres)
572{
573 struct inode *inode = ocfs2_lock_res_inode(lockres);
574
575 return OCFS2_SB(inode->i_sb);
576}
577
578static struct ocfs2_super *ocfs2_get_qinfo_osb(struct ocfs2_lock_res *lockres)
579{
580 struct ocfs2_mem_dqinfo *info = lockres->l_priv;
581
582 return OCFS2_SB(info->dqi_gi.dqi_sb);
583}
584
585static struct ocfs2_super *ocfs2_get_file_osb(struct ocfs2_lock_res *lockres)
586{
587 struct ocfs2_file_private *fp = lockres->l_priv;
588
589 return OCFS2_SB(fp->fp_file->f_mapping->host->i_sb);
590}
591
592static __u64 ocfs2_get_dentry_lock_ino(struct ocfs2_lock_res *lockres)
593{
594 __be64 inode_blkno_be;
595
596 memcpy(&inode_blkno_be, &lockres->l_name[OCFS2_DENTRY_LOCK_INO_START],
597 sizeof(__be64));
598
599 return be64_to_cpu(inode_blkno_be);
600}
601
602static struct ocfs2_super *ocfs2_get_dentry_osb(struct ocfs2_lock_res *lockres)
603{
604 struct ocfs2_dentry_lock *dl = lockres->l_priv;
605
606 return OCFS2_SB(dl->dl_inode->i_sb);
607}
608
609void ocfs2_dentry_lock_res_init(struct ocfs2_dentry_lock *dl,
610 u64 parent, struct inode *inode)
611{
612 int len;
613 u64 inode_blkno = OCFS2_I(inode)->ip_blkno;
614 __be64 inode_blkno_be = cpu_to_be64(inode_blkno);
615 struct ocfs2_lock_res *lockres = &dl->dl_lockres;
616
617 ocfs2_lock_res_init_once(lockres);
618
619 /*
620 * Unfortunately, the standard lock naming scheme won't work
621 * here because we have two 16 byte values to use. Instead,
622 * we'll stuff the inode number as a binary value. We still
623 * want error prints to show something without garbling the
624 * display, so drop a null byte in there before the inode
625 * number. A future version of OCFS2 will likely use all
626 * binary lock names. The stringified names have been a
627 * tremendous aid in debugging, but now that the debugfs
628 * interface exists, we can mangle things there if need be.
629 *
630 * NOTE: We also drop the standard "pad" value (the total lock
631 * name size stays the same though - the last part is all
632 * zeros due to the memset in ocfs2_lock_res_init_once()
633 */
634 len = snprintf(lockres->l_name, OCFS2_DENTRY_LOCK_INO_START,
635 "%c%016llx",
636 ocfs2_lock_type_char(OCFS2_LOCK_TYPE_DENTRY),
637 (long long)parent);
638
639 BUG_ON(len != (OCFS2_DENTRY_LOCK_INO_START - 1));
640
641 memcpy(&lockres->l_name[OCFS2_DENTRY_LOCK_INO_START], &inode_blkno_be,
642 sizeof(__be64));
643
644 ocfs2_lock_res_init_common(OCFS2_SB(inode->i_sb), lockres,
645 OCFS2_LOCK_TYPE_DENTRY, &ocfs2_dentry_lops,
646 dl);
647}
648
649static void ocfs2_super_lock_res_init(struct ocfs2_lock_res *res,
650 struct ocfs2_super *osb)
651{
652 /* Superblock lockres doesn't come from a slab so we call init
653 * once on it manually. */
654 ocfs2_lock_res_init_once(res);
655 ocfs2_build_lock_name(OCFS2_LOCK_TYPE_SUPER, OCFS2_SUPER_BLOCK_BLKNO,
656 0, res->l_name);
657 ocfs2_lock_res_init_common(osb, res, OCFS2_LOCK_TYPE_SUPER,
658 &ocfs2_super_lops, osb);
659}
660
661static void ocfs2_rename_lock_res_init(struct ocfs2_lock_res *res,
662 struct ocfs2_super *osb)
663{
664 /* Rename lockres doesn't come from a slab so we call init
665 * once on it manually. */
666 ocfs2_lock_res_init_once(res);
667 ocfs2_build_lock_name(OCFS2_LOCK_TYPE_RENAME, 0, 0, res->l_name);
668 ocfs2_lock_res_init_common(osb, res, OCFS2_LOCK_TYPE_RENAME,
669 &ocfs2_rename_lops, osb);
670}
671
672static void ocfs2_nfs_sync_lock_res_init(struct ocfs2_lock_res *res,
673 struct ocfs2_super *osb)
674{
675 /* nfs_sync lockres doesn't come from a slab so we call init
676 * once on it manually. */
677 ocfs2_lock_res_init_once(res);
678 ocfs2_build_lock_name(OCFS2_LOCK_TYPE_NFS_SYNC, 0, 0, res->l_name);
679 ocfs2_lock_res_init_common(osb, res, OCFS2_LOCK_TYPE_NFS_SYNC,
680 &ocfs2_nfs_sync_lops, osb);
681}
682
683void ocfs2_trim_fs_lock_res_init(struct ocfs2_super *osb)
684{
685 struct ocfs2_lock_res *lockres = &osb->osb_trim_fs_lockres;
686
687 ocfs2_lock_res_init_once(lockres);
688 ocfs2_build_lock_name(OCFS2_LOCK_TYPE_TRIM_FS, 0, 0, lockres->l_name);
689 ocfs2_lock_res_init_common(osb, lockres, OCFS2_LOCK_TYPE_TRIM_FS,
690 &ocfs2_trim_fs_lops, osb);
691}
692
693void ocfs2_trim_fs_lock_res_uninit(struct ocfs2_super *osb)
694{
695 struct ocfs2_lock_res *lockres = &osb->osb_trim_fs_lockres;
696
697 ocfs2_simple_drop_lockres(osb, lockres);
698 ocfs2_lock_res_free(lockres);
699}
700
701static void ocfs2_orphan_scan_lock_res_init(struct ocfs2_lock_res *res,
702 struct ocfs2_super *osb)
703{
704 ocfs2_lock_res_init_once(res);
705 ocfs2_build_lock_name(OCFS2_LOCK_TYPE_ORPHAN_SCAN, 0, 0, res->l_name);
706 ocfs2_lock_res_init_common(osb, res, OCFS2_LOCK_TYPE_ORPHAN_SCAN,
707 &ocfs2_orphan_scan_lops, osb);
708}
709
710void ocfs2_file_lock_res_init(struct ocfs2_lock_res *lockres,
711 struct ocfs2_file_private *fp)
712{
713 struct inode *inode = fp->fp_file->f_mapping->host;
714 struct ocfs2_inode_info *oi = OCFS2_I(inode);
715
716 ocfs2_lock_res_init_once(lockres);
717 ocfs2_build_lock_name(OCFS2_LOCK_TYPE_FLOCK, oi->ip_blkno,
718 inode->i_generation, lockres->l_name);
719 ocfs2_lock_res_init_common(OCFS2_SB(inode->i_sb), lockres,
720 OCFS2_LOCK_TYPE_FLOCK, &ocfs2_flock_lops,
721 fp);
722 lockres->l_flags |= OCFS2_LOCK_NOCACHE;
723}
724
725void ocfs2_qinfo_lock_res_init(struct ocfs2_lock_res *lockres,
726 struct ocfs2_mem_dqinfo *info)
727{
728 ocfs2_lock_res_init_once(lockres);
729 ocfs2_build_lock_name(OCFS2_LOCK_TYPE_QINFO, info->dqi_gi.dqi_type,
730 0, lockres->l_name);
731 ocfs2_lock_res_init_common(OCFS2_SB(info->dqi_gi.dqi_sb), lockres,
732 OCFS2_LOCK_TYPE_QINFO, &ocfs2_qinfo_lops,
733 info);
734}
735
736void ocfs2_refcount_lock_res_init(struct ocfs2_lock_res *lockres,
737 struct ocfs2_super *osb, u64 ref_blkno,
738 unsigned int generation)
739{
740 ocfs2_lock_res_init_once(lockres);
741 ocfs2_build_lock_name(OCFS2_LOCK_TYPE_REFCOUNT, ref_blkno,
742 generation, lockres->l_name);
743 ocfs2_lock_res_init_common(osb, lockres, OCFS2_LOCK_TYPE_REFCOUNT,
744 &ocfs2_refcount_block_lops, osb);
745}
746
747void ocfs2_lock_res_free(struct ocfs2_lock_res *res)
748{
749 if (!(res->l_flags & OCFS2_LOCK_INITIALIZED))
750 return;
751
752 ocfs2_remove_lockres_tracking(res);
753
754 mlog_bug_on_msg(!list_empty(&res->l_blocked_list),
755 "Lockres %s is on the blocked list\n",
756 res->l_name);
757 mlog_bug_on_msg(!list_empty(&res->l_mask_waiters),
758 "Lockres %s has mask waiters pending\n",
759 res->l_name);
760 mlog_bug_on_msg(spin_is_locked(&res->l_lock),
761 "Lockres %s is locked\n",
762 res->l_name);
763 mlog_bug_on_msg(res->l_ro_holders,
764 "Lockres %s has %u ro holders\n",
765 res->l_name, res->l_ro_holders);
766 mlog_bug_on_msg(res->l_ex_holders,
767 "Lockres %s has %u ex holders\n",
768 res->l_name, res->l_ex_holders);
769
770 /* Need to clear out the lock status block for the dlm */
771 memset(&res->l_lksb, 0, sizeof(res->l_lksb));
772
773 res->l_flags = 0UL;
774}
775
776/*
777 * Keep a list of processes who have interest in a lockres.
778 * Note: this is now only uesed for check recursive cluster locking.
779 */
780static inline void ocfs2_add_holder(struct ocfs2_lock_res *lockres,
781 struct ocfs2_lock_holder *oh)
782{
783 INIT_LIST_HEAD(&oh->oh_list);
784 oh->oh_owner_pid = get_pid(task_pid(current));
785
786 spin_lock(&lockres->l_lock);
787 list_add_tail(&oh->oh_list, &lockres->l_holders);
788 spin_unlock(&lockres->l_lock);
789}
790
791static inline void ocfs2_remove_holder(struct ocfs2_lock_res *lockres,
792 struct ocfs2_lock_holder *oh)
793{
794 spin_lock(&lockres->l_lock);
795 list_del(&oh->oh_list);
796 spin_unlock(&lockres->l_lock);
797
798 put_pid(oh->oh_owner_pid);
799}
800
801static inline int ocfs2_is_locked_by_me(struct ocfs2_lock_res *lockres)
802{
803 struct ocfs2_lock_holder *oh;
804 struct pid *pid;
805
806 /* look in the list of holders for one with the current task as owner */
807 spin_lock(&lockres->l_lock);
808 pid = task_pid(current);
809 list_for_each_entry(oh, &lockres->l_holders, oh_list) {
810 if (oh->oh_owner_pid == pid) {
811 spin_unlock(&lockres->l_lock);
812 return 1;
813 }
814 }
815 spin_unlock(&lockres->l_lock);
816
817 return 0;
818}
819
820static inline void ocfs2_inc_holders(struct ocfs2_lock_res *lockres,
821 int level)
822{
823 BUG_ON(!lockres);
824
825 switch(level) {
826 case DLM_LOCK_EX:
827 lockres->l_ex_holders++;
828 break;
829 case DLM_LOCK_PR:
830 lockres->l_ro_holders++;
831 break;
832 default:
833 BUG();
834 }
835}
836
837static inline void ocfs2_dec_holders(struct ocfs2_lock_res *lockres,
838 int level)
839{
840 BUG_ON(!lockres);
841
842 switch(level) {
843 case DLM_LOCK_EX:
844 BUG_ON(!lockres->l_ex_holders);
845 lockres->l_ex_holders--;
846 break;
847 case DLM_LOCK_PR:
848 BUG_ON(!lockres->l_ro_holders);
849 lockres->l_ro_holders--;
850 break;
851 default:
852 BUG();
853 }
854}
855
856/* WARNING: This function lives in a world where the only three lock
857 * levels are EX, PR, and NL. It *will* have to be adjusted when more
858 * lock types are added. */
859static inline int ocfs2_highest_compat_lock_level(int level)
860{
861 int new_level = DLM_LOCK_EX;
862
863 if (level == DLM_LOCK_EX)
864 new_level = DLM_LOCK_NL;
865 else if (level == DLM_LOCK_PR)
866 new_level = DLM_LOCK_PR;
867 return new_level;
868}
869
870static void lockres_set_flags(struct ocfs2_lock_res *lockres,
871 unsigned long newflags)
872{
873 struct ocfs2_mask_waiter *mw, *tmp;
874
875 assert_spin_locked(&lockres->l_lock);
876
877 lockres->l_flags = newflags;
878
879 list_for_each_entry_safe(mw, tmp, &lockres->l_mask_waiters, mw_item) {
880 if ((lockres->l_flags & mw->mw_mask) != mw->mw_goal)
881 continue;
882
883 list_del_init(&mw->mw_item);
884 mw->mw_status = 0;
885 complete(&mw->mw_complete);
886 }
887}
888static void lockres_or_flags(struct ocfs2_lock_res *lockres, unsigned long or)
889{
890 lockres_set_flags(lockres, lockres->l_flags | or);
891}
892static void lockres_clear_flags(struct ocfs2_lock_res *lockres,
893 unsigned long clear)
894{
895 lockres_set_flags(lockres, lockres->l_flags & ~clear);
896}
897
898static inline void ocfs2_generic_handle_downconvert_action(struct ocfs2_lock_res *lockres)
899{
900 BUG_ON(!(lockres->l_flags & OCFS2_LOCK_BUSY));
901 BUG_ON(!(lockres->l_flags & OCFS2_LOCK_ATTACHED));
902 BUG_ON(!(lockres->l_flags & OCFS2_LOCK_BLOCKED));
903 BUG_ON(lockres->l_blocking <= DLM_LOCK_NL);
904
905 lockres->l_level = lockres->l_requested;
906 if (lockres->l_level <=
907 ocfs2_highest_compat_lock_level(lockres->l_blocking)) {
908 lockres->l_blocking = DLM_LOCK_NL;
909 lockres_clear_flags(lockres, OCFS2_LOCK_BLOCKED);
910 }
911 lockres_clear_flags(lockres, OCFS2_LOCK_BUSY);
912}
913
914static inline void ocfs2_generic_handle_convert_action(struct ocfs2_lock_res *lockres)
915{
916 BUG_ON(!(lockres->l_flags & OCFS2_LOCK_BUSY));
917 BUG_ON(!(lockres->l_flags & OCFS2_LOCK_ATTACHED));
918
919 /* Convert from RO to EX doesn't really need anything as our
920 * information is already up to data. Convert from NL to
921 * *anything* however should mark ourselves as needing an
922 * update */
923 if (lockres->l_level == DLM_LOCK_NL &&
924 lockres->l_ops->flags & LOCK_TYPE_REQUIRES_REFRESH)
925 lockres_or_flags(lockres, OCFS2_LOCK_NEEDS_REFRESH);
926
927 lockres->l_level = lockres->l_requested;
928
929 /*
930 * We set the OCFS2_LOCK_UPCONVERT_FINISHING flag before clearing
931 * the OCFS2_LOCK_BUSY flag to prevent the dc thread from
932 * downconverting the lock before the upconvert has fully completed.
933 * Do not prevent the dc thread from downconverting if NONBLOCK lock
934 * had already returned.
935 */
936 if (!(lockres->l_flags & OCFS2_LOCK_NONBLOCK_FINISHED))
937 lockres_or_flags(lockres, OCFS2_LOCK_UPCONVERT_FINISHING);
938 else
939 lockres_clear_flags(lockres, OCFS2_LOCK_NONBLOCK_FINISHED);
940
941 lockres_clear_flags(lockres, OCFS2_LOCK_BUSY);
942}
943
944static inline void ocfs2_generic_handle_attach_action(struct ocfs2_lock_res *lockres)
945{
946 BUG_ON((!(lockres->l_flags & OCFS2_LOCK_BUSY)));
947 BUG_ON(lockres->l_flags & OCFS2_LOCK_ATTACHED);
948
949 if (lockres->l_requested > DLM_LOCK_NL &&
950 !(lockres->l_flags & OCFS2_LOCK_LOCAL) &&
951 lockres->l_ops->flags & LOCK_TYPE_REQUIRES_REFRESH)
952 lockres_or_flags(lockres, OCFS2_LOCK_NEEDS_REFRESH);
953
954 lockres->l_level = lockres->l_requested;
955 lockres_or_flags(lockres, OCFS2_LOCK_ATTACHED);
956 lockres_clear_flags(lockres, OCFS2_LOCK_BUSY);
957}
958
959static int ocfs2_generic_handle_bast(struct ocfs2_lock_res *lockres,
960 int level)
961{
962 int needs_downconvert = 0;
963
964 assert_spin_locked(&lockres->l_lock);
965
966 if (level > lockres->l_blocking) {
967 /* only schedule a downconvert if we haven't already scheduled
968 * one that goes low enough to satisfy the level we're
969 * blocking. this also catches the case where we get
970 * duplicate BASTs */
971 if (ocfs2_highest_compat_lock_level(level) <
972 ocfs2_highest_compat_lock_level(lockres->l_blocking))
973 needs_downconvert = 1;
974
975 lockres->l_blocking = level;
976 }
977
978 mlog(ML_BASTS, "lockres %s, block %d, level %d, l_block %d, dwn %d\n",
979 lockres->l_name, level, lockres->l_level, lockres->l_blocking,
980 needs_downconvert);
981
982 if (needs_downconvert)
983 lockres_or_flags(lockres, OCFS2_LOCK_BLOCKED);
984 mlog(0, "needs_downconvert = %d\n", needs_downconvert);
985 return needs_downconvert;
986}
987
988/*
989 * OCFS2_LOCK_PENDING and l_pending_gen.
990 *
991 * Why does OCFS2_LOCK_PENDING exist? To close a race between setting
992 * OCFS2_LOCK_BUSY and calling ocfs2_dlm_lock(). See ocfs2_unblock_lock()
993 * for more details on the race.
994 *
995 * OCFS2_LOCK_PENDING closes the race quite nicely. However, it introduces
996 * a race on itself. In o2dlm, we can get the ast before ocfs2_dlm_lock()
997 * returns. The ast clears OCFS2_LOCK_BUSY, and must therefore clear
998 * OCFS2_LOCK_PENDING at the same time. When ocfs2_dlm_lock() returns,
999 * the caller is going to try to clear PENDING again. If nothing else is
1000 * happening, __lockres_clear_pending() sees PENDING is unset and does
1001 * nothing.
1002 *
1003 * But what if another path (eg downconvert thread) has just started a
1004 * new locking action? The other path has re-set PENDING. Our path
1005 * cannot clear PENDING, because that will re-open the original race
1006 * window.
1007 *
1008 * [Example]
1009 *
1010 * ocfs2_meta_lock()
1011 * ocfs2_cluster_lock()
1012 * set BUSY
1013 * set PENDING
1014 * drop l_lock
1015 * ocfs2_dlm_lock()
1016 * ocfs2_locking_ast() ocfs2_downconvert_thread()
1017 * clear PENDING ocfs2_unblock_lock()
1018 * take_l_lock
1019 * !BUSY
1020 * ocfs2_prepare_downconvert()
1021 * set BUSY
1022 * set PENDING
1023 * drop l_lock
1024 * take l_lock
1025 * clear PENDING
1026 * drop l_lock
1027 * <window>
1028 * ocfs2_dlm_lock()
1029 *
1030 * So as you can see, we now have a window where l_lock is not held,
1031 * PENDING is not set, and ocfs2_dlm_lock() has not been called.
1032 *
1033 * The core problem is that ocfs2_cluster_lock() has cleared the PENDING
1034 * set by ocfs2_prepare_downconvert(). That wasn't nice.
1035 *
1036 * To solve this we introduce l_pending_gen. A call to
1037 * lockres_clear_pending() will only do so when it is passed a generation
1038 * number that matches the lockres. lockres_set_pending() will return the
1039 * current generation number. When ocfs2_cluster_lock() goes to clear
1040 * PENDING, it passes the generation it got from set_pending(). In our
1041 * example above, the generation numbers will *not* match. Thus,
1042 * ocfs2_cluster_lock() will not clear the PENDING set by
1043 * ocfs2_prepare_downconvert().
1044 */
1045
1046/* Unlocked version for ocfs2_locking_ast() */
1047static void __lockres_clear_pending(struct ocfs2_lock_res *lockres,
1048 unsigned int generation,
1049 struct ocfs2_super *osb)
1050{
1051 assert_spin_locked(&lockres->l_lock);
1052
1053 /*
1054 * The ast and locking functions can race us here. The winner
1055 * will clear pending, the loser will not.
1056 */
1057 if (!(lockres->l_flags & OCFS2_LOCK_PENDING) ||
1058 (lockres->l_pending_gen != generation))
1059 return;
1060
1061 lockres_clear_flags(lockres, OCFS2_LOCK_PENDING);
1062 lockres->l_pending_gen++;
1063
1064 /*
1065 * The downconvert thread may have skipped us because we
1066 * were PENDING. Wake it up.
1067 */
1068 if (lockres->l_flags & OCFS2_LOCK_BLOCKED)
1069 ocfs2_wake_downconvert_thread(osb);
1070}
1071
1072/* Locked version for callers of ocfs2_dlm_lock() */
1073static void lockres_clear_pending(struct ocfs2_lock_res *lockres,
1074 unsigned int generation,
1075 struct ocfs2_super *osb)
1076{
1077 unsigned long flags;
1078
1079 spin_lock_irqsave(&lockres->l_lock, flags);
1080 __lockres_clear_pending(lockres, generation, osb);
1081 spin_unlock_irqrestore(&lockres->l_lock, flags);
1082}
1083
1084static unsigned int lockres_set_pending(struct ocfs2_lock_res *lockres)
1085{
1086 assert_spin_locked(&lockres->l_lock);
1087 BUG_ON(!(lockres->l_flags & OCFS2_LOCK_BUSY));
1088
1089 lockres_or_flags(lockres, OCFS2_LOCK_PENDING);
1090
1091 return lockres->l_pending_gen;
1092}
1093
1094static void ocfs2_blocking_ast(struct ocfs2_dlm_lksb *lksb, int level)
1095{
1096 struct ocfs2_lock_res *lockres = ocfs2_lksb_to_lock_res(lksb);
1097 struct ocfs2_super *osb = ocfs2_get_lockres_osb(lockres);
1098 int needs_downconvert;
1099 unsigned long flags;
1100
1101 BUG_ON(level <= DLM_LOCK_NL);
1102
1103 mlog(ML_BASTS, "BAST fired for lockres %s, blocking %d, level %d, "
1104 "type %s\n", lockres->l_name, level, lockres->l_level,
1105 ocfs2_lock_type_string(lockres->l_type));
1106
1107 /*
1108 * We can skip the bast for locks which don't enable caching -
1109 * they'll be dropped at the earliest possible time anyway.
1110 */
1111 if (lockres->l_flags & OCFS2_LOCK_NOCACHE)
1112 return;
1113
1114 spin_lock_irqsave(&lockres->l_lock, flags);
1115 needs_downconvert = ocfs2_generic_handle_bast(lockres, level);
1116 if (needs_downconvert)
1117 ocfs2_schedule_blocked_lock(osb, lockres);
1118 spin_unlock_irqrestore(&lockres->l_lock, flags);
1119
1120 wake_up(&lockres->l_event);
1121
1122 ocfs2_wake_downconvert_thread(osb);
1123}
1124
1125static void ocfs2_locking_ast(struct ocfs2_dlm_lksb *lksb)
1126{
1127 struct ocfs2_lock_res *lockres = ocfs2_lksb_to_lock_res(lksb);
1128 struct ocfs2_super *osb = ocfs2_get_lockres_osb(lockres);
1129 unsigned long flags;
1130 int status;
1131
1132 spin_lock_irqsave(&lockres->l_lock, flags);
1133
1134 status = ocfs2_dlm_lock_status(&lockres->l_lksb);
1135
1136 if (status == -EAGAIN) {
1137 lockres_clear_flags(lockres, OCFS2_LOCK_BUSY);
1138 goto out;
1139 }
1140
1141 if (status) {
1142 mlog(ML_ERROR, "lockres %s: lksb status value of %d!\n",
1143 lockres->l_name, status);
1144 spin_unlock_irqrestore(&lockres->l_lock, flags);
1145 return;
1146 }
1147
1148 mlog(ML_BASTS, "AST fired for lockres %s, action %d, unlock %d, "
1149 "level %d => %d\n", lockres->l_name, lockres->l_action,
1150 lockres->l_unlock_action, lockres->l_level, lockres->l_requested);
1151
1152 switch(lockres->l_action) {
1153 case OCFS2_AST_ATTACH:
1154 ocfs2_generic_handle_attach_action(lockres);
1155 lockres_clear_flags(lockres, OCFS2_LOCK_LOCAL);
1156 break;
1157 case OCFS2_AST_CONVERT:
1158 ocfs2_generic_handle_convert_action(lockres);
1159 break;
1160 case OCFS2_AST_DOWNCONVERT:
1161 ocfs2_generic_handle_downconvert_action(lockres);
1162 break;
1163 default:
1164 mlog(ML_ERROR, "lockres %s: AST fired with invalid action: %u, "
1165 "flags 0x%lx, unlock: %u\n",
1166 lockres->l_name, lockres->l_action, lockres->l_flags,
1167 lockres->l_unlock_action);
1168 BUG();
1169 }
1170out:
1171 /* set it to something invalid so if we get called again we
1172 * can catch it. */
1173 lockres->l_action = OCFS2_AST_INVALID;
1174
1175 /* Did we try to cancel this lock? Clear that state */
1176 if (lockres->l_unlock_action == OCFS2_UNLOCK_CANCEL_CONVERT)
1177 lockres->l_unlock_action = OCFS2_UNLOCK_INVALID;
1178
1179 /*
1180 * We may have beaten the locking functions here. We certainly
1181 * know that dlm_lock() has been called :-)
1182 * Because we can't have two lock calls in flight at once, we
1183 * can use lockres->l_pending_gen.
1184 */
1185 __lockres_clear_pending(lockres, lockres->l_pending_gen, osb);
1186
1187 wake_up(&lockres->l_event);
1188 spin_unlock_irqrestore(&lockres->l_lock, flags);
1189}
1190
1191static void ocfs2_unlock_ast(struct ocfs2_dlm_lksb *lksb, int error)
1192{
1193 struct ocfs2_lock_res *lockres = ocfs2_lksb_to_lock_res(lksb);
1194 unsigned long flags;
1195
1196 mlog(ML_BASTS, "UNLOCK AST fired for lockres %s, action = %d\n",
1197 lockres->l_name, lockres->l_unlock_action);
1198
1199 spin_lock_irqsave(&lockres->l_lock, flags);
1200 if (error) {
1201 mlog(ML_ERROR, "Dlm passes error %d for lock %s, "
1202 "unlock_action %d\n", error, lockres->l_name,
1203 lockres->l_unlock_action);
1204 spin_unlock_irqrestore(&lockres->l_lock, flags);
1205 return;
1206 }
1207
1208 switch(lockres->l_unlock_action) {
1209 case OCFS2_UNLOCK_CANCEL_CONVERT:
1210 mlog(0, "Cancel convert success for %s\n", lockres->l_name);
1211 lockres->l_action = OCFS2_AST_INVALID;
1212 /* Downconvert thread may have requeued this lock, we
1213 * need to wake it. */
1214 if (lockres->l_flags & OCFS2_LOCK_BLOCKED)
1215 ocfs2_wake_downconvert_thread(ocfs2_get_lockres_osb(lockres));
1216 break;
1217 case OCFS2_UNLOCK_DROP_LOCK:
1218 lockres->l_level = DLM_LOCK_IV;
1219 break;
1220 default:
1221 BUG();
1222 }
1223
1224 lockres_clear_flags(lockres, OCFS2_LOCK_BUSY);
1225 lockres->l_unlock_action = OCFS2_UNLOCK_INVALID;
1226 wake_up(&lockres->l_event);
1227 spin_unlock_irqrestore(&lockres->l_lock, flags);
1228}
1229
1230/*
1231 * This is the filesystem locking protocol. It provides the lock handling
1232 * hooks for the underlying DLM. It has a maximum version number.
1233 * The version number allows interoperability with systems running at
1234 * the same major number and an equal or smaller minor number.
1235 *
1236 * Whenever the filesystem does new things with locks (adds or removes a
1237 * lock, orders them differently, does different things underneath a lock),
1238 * the version must be changed. The protocol is negotiated when joining
1239 * the dlm domain. A node may join the domain if its major version is
1240 * identical to all other nodes and its minor version is greater than
1241 * or equal to all other nodes. When its minor version is greater than
1242 * the other nodes, it will run at the minor version specified by the
1243 * other nodes.
1244 *
1245 * If a locking change is made that will not be compatible with older
1246 * versions, the major number must be increased and the minor version set
1247 * to zero. If a change merely adds a behavior that can be disabled when
1248 * speaking to older versions, the minor version must be increased. If a
1249 * change adds a fully backwards compatible change (eg, LVB changes that
1250 * are just ignored by older versions), the version does not need to be
1251 * updated.
1252 */
1253static struct ocfs2_locking_protocol lproto = {
1254 .lp_max_version = {
1255 .pv_major = OCFS2_LOCKING_PROTOCOL_MAJOR,
1256 .pv_minor = OCFS2_LOCKING_PROTOCOL_MINOR,
1257 },
1258 .lp_lock_ast = ocfs2_locking_ast,
1259 .lp_blocking_ast = ocfs2_blocking_ast,
1260 .lp_unlock_ast = ocfs2_unlock_ast,
1261};
1262
1263void ocfs2_set_locking_protocol(void)
1264{
1265 ocfs2_stack_glue_set_max_proto_version(&lproto.lp_max_version);
1266}
1267
1268static inline void ocfs2_recover_from_dlm_error(struct ocfs2_lock_res *lockres,
1269 int convert)
1270{
1271 unsigned long flags;
1272
1273 spin_lock_irqsave(&lockres->l_lock, flags);
1274 lockres_clear_flags(lockres, OCFS2_LOCK_BUSY);
1275 lockres_clear_flags(lockres, OCFS2_LOCK_UPCONVERT_FINISHING);
1276 if (convert)
1277 lockres->l_action = OCFS2_AST_INVALID;
1278 else
1279 lockres->l_unlock_action = OCFS2_UNLOCK_INVALID;
1280 spin_unlock_irqrestore(&lockres->l_lock, flags);
1281
1282 wake_up(&lockres->l_event);
1283}
1284
1285/* Note: If we detect another process working on the lock (i.e.,
1286 * OCFS2_LOCK_BUSY), we'll bail out returning 0. It's up to the caller
1287 * to do the right thing in that case.
1288 */
1289static int ocfs2_lock_create(struct ocfs2_super *osb,
1290 struct ocfs2_lock_res *lockres,
1291 int level,
1292 u32 dlm_flags)
1293{
1294 int ret = 0;
1295 unsigned long flags;
1296 unsigned int gen;
1297
1298 mlog(0, "lock %s, level = %d, flags = %u\n", lockres->l_name, level,
1299 dlm_flags);
1300
1301 spin_lock_irqsave(&lockres->l_lock, flags);
1302 if ((lockres->l_flags & OCFS2_LOCK_ATTACHED) ||
1303 (lockres->l_flags & OCFS2_LOCK_BUSY)) {
1304 spin_unlock_irqrestore(&lockres->l_lock, flags);
1305 goto bail;
1306 }
1307
1308 lockres->l_action = OCFS2_AST_ATTACH;
1309 lockres->l_requested = level;
1310 lockres_or_flags(lockres, OCFS2_LOCK_BUSY);
1311 gen = lockres_set_pending(lockres);
1312 spin_unlock_irqrestore(&lockres->l_lock, flags);
1313
1314 ret = ocfs2_dlm_lock(osb->cconn,
1315 level,
1316 &lockres->l_lksb,
1317 dlm_flags,
1318 lockres->l_name,
1319 OCFS2_LOCK_ID_MAX_LEN - 1);
1320 lockres_clear_pending(lockres, gen, osb);
1321 if (ret) {
1322 ocfs2_log_dlm_error("ocfs2_dlm_lock", ret, lockres);
1323 ocfs2_recover_from_dlm_error(lockres, 1);
1324 }
1325
1326 mlog(0, "lock %s, return from ocfs2_dlm_lock\n", lockres->l_name);
1327
1328bail:
1329 return ret;
1330}
1331
1332static inline int ocfs2_check_wait_flag(struct ocfs2_lock_res *lockres,
1333 int flag)
1334{
1335 unsigned long flags;
1336 int ret;
1337
1338 spin_lock_irqsave(&lockres->l_lock, flags);
1339 ret = lockres->l_flags & flag;
1340 spin_unlock_irqrestore(&lockres->l_lock, flags);
1341
1342 return ret;
1343}
1344
1345static inline void ocfs2_wait_on_busy_lock(struct ocfs2_lock_res *lockres)
1346
1347{
1348 wait_event(lockres->l_event,
1349 !ocfs2_check_wait_flag(lockres, OCFS2_LOCK_BUSY));
1350}
1351
1352static inline void ocfs2_wait_on_refreshing_lock(struct ocfs2_lock_res *lockres)
1353
1354{
1355 wait_event(lockres->l_event,
1356 !ocfs2_check_wait_flag(lockres, OCFS2_LOCK_REFRESHING));
1357}
1358
1359/* predict what lock level we'll be dropping down to on behalf
1360 * of another node, and return true if the currently wanted
1361 * level will be compatible with it. */
1362static inline int ocfs2_may_continue_on_blocked_lock(struct ocfs2_lock_res *lockres,
1363 int wanted)
1364{
1365 BUG_ON(!(lockres->l_flags & OCFS2_LOCK_BLOCKED));
1366
1367 return wanted <= ocfs2_highest_compat_lock_level(lockres->l_blocking);
1368}
1369
1370static void ocfs2_init_mask_waiter(struct ocfs2_mask_waiter *mw)
1371{
1372 INIT_LIST_HEAD(&mw->mw_item);
1373 init_completion(&mw->mw_complete);
1374 ocfs2_init_start_time(mw);
1375}
1376
1377static int ocfs2_wait_for_mask(struct ocfs2_mask_waiter *mw)
1378{
1379 wait_for_completion(&mw->mw_complete);
1380 /* Re-arm the completion in case we want to wait on it again */
1381 reinit_completion(&mw->mw_complete);
1382 return mw->mw_status;
1383}
1384
1385static void lockres_add_mask_waiter(struct ocfs2_lock_res *lockres,
1386 struct ocfs2_mask_waiter *mw,
1387 unsigned long mask,
1388 unsigned long goal)
1389{
1390 BUG_ON(!list_empty(&mw->mw_item));
1391
1392 assert_spin_locked(&lockres->l_lock);
1393
1394 list_add_tail(&mw->mw_item, &lockres->l_mask_waiters);
1395 mw->mw_mask = mask;
1396 mw->mw_goal = goal;
1397}
1398
1399/* returns 0 if the mw that was removed was already satisfied, -EBUSY
1400 * if the mask still hadn't reached its goal */
1401static int __lockres_remove_mask_waiter(struct ocfs2_lock_res *lockres,
1402 struct ocfs2_mask_waiter *mw)
1403{
1404 int ret = 0;
1405
1406 assert_spin_locked(&lockres->l_lock);
1407 if (!list_empty(&mw->mw_item)) {
1408 if ((lockres->l_flags & mw->mw_mask) != mw->mw_goal)
1409 ret = -EBUSY;
1410
1411 list_del_init(&mw->mw_item);
1412 init_completion(&mw->mw_complete);
1413 }
1414
1415 return ret;
1416}
1417
1418static int lockres_remove_mask_waiter(struct ocfs2_lock_res *lockres,
1419 struct ocfs2_mask_waiter *mw)
1420{
1421 unsigned long flags;
1422 int ret = 0;
1423
1424 spin_lock_irqsave(&lockres->l_lock, flags);
1425 ret = __lockres_remove_mask_waiter(lockres, mw);
1426 spin_unlock_irqrestore(&lockres->l_lock, flags);
1427
1428 return ret;
1429
1430}
1431
1432static int ocfs2_wait_for_mask_interruptible(struct ocfs2_mask_waiter *mw,
1433 struct ocfs2_lock_res *lockres)
1434{
1435 int ret;
1436
1437 ret = wait_for_completion_interruptible(&mw->mw_complete);
1438 if (ret)
1439 lockres_remove_mask_waiter(lockres, mw);
1440 else
1441 ret = mw->mw_status;
1442 /* Re-arm the completion in case we want to wait on it again */
1443 reinit_completion(&mw->mw_complete);
1444 return ret;
1445}
1446
1447static int __ocfs2_cluster_lock(struct ocfs2_super *osb,
1448 struct ocfs2_lock_res *lockres,
1449 int level,
1450 u32 lkm_flags,
1451 int arg_flags,
1452 int l_subclass,
1453 unsigned long caller_ip)
1454{
1455 struct ocfs2_mask_waiter mw;
1456 int wait, catch_signals = !(osb->s_mount_opt & OCFS2_MOUNT_NOINTR);
1457 int ret = 0; /* gcc doesn't realize wait = 1 guarantees ret is set */
1458 unsigned long flags;
1459 unsigned int gen;
1460 int noqueue_attempted = 0;
1461 int dlm_locked = 0;
1462 int kick_dc = 0;
1463
1464 if (!(lockres->l_flags & OCFS2_LOCK_INITIALIZED)) {
1465 mlog_errno(-EINVAL);
1466 return -EINVAL;
1467 }
1468
1469 ocfs2_init_mask_waiter(&mw);
1470
1471 if (lockres->l_ops->flags & LOCK_TYPE_USES_LVB)
1472 lkm_flags |= DLM_LKF_VALBLK;
1473
1474again:
1475 wait = 0;
1476
1477 spin_lock_irqsave(&lockres->l_lock, flags);
1478
1479 if (catch_signals && signal_pending(current)) {
1480 ret = -ERESTARTSYS;
1481 goto unlock;
1482 }
1483
1484 mlog_bug_on_msg(lockres->l_flags & OCFS2_LOCK_FREEING,
1485 "Cluster lock called on freeing lockres %s! flags "
1486 "0x%lx\n", lockres->l_name, lockres->l_flags);
1487
1488 /* We only compare against the currently granted level
1489 * here. If the lock is blocked waiting on a downconvert,
1490 * we'll get caught below. */
1491 if (lockres->l_flags & OCFS2_LOCK_BUSY &&
1492 level > lockres->l_level) {
1493 /* is someone sitting in dlm_lock? If so, wait on
1494 * them. */
1495 lockres_add_mask_waiter(lockres, &mw, OCFS2_LOCK_BUSY, 0);
1496 wait = 1;
1497 goto unlock;
1498 }
1499
1500 if (lockres->l_flags & OCFS2_LOCK_UPCONVERT_FINISHING) {
1501 /*
1502 * We've upconverted. If the lock now has a level we can
1503 * work with, we take it. If, however, the lock is not at the
1504 * required level, we go thru the full cycle. One way this could
1505 * happen is if a process requesting an upconvert to PR is
1506 * closely followed by another requesting upconvert to an EX.
1507 * If the process requesting EX lands here, we want it to
1508 * continue attempting to upconvert and let the process
1509 * requesting PR take the lock.
1510 * If multiple processes request upconvert to PR, the first one
1511 * here will take the lock. The others will have to go thru the
1512 * OCFS2_LOCK_BLOCKED check to ensure that there is no pending
1513 * downconvert request.
1514 */
1515 if (level <= lockres->l_level)
1516 goto update_holders;
1517 }
1518
1519 if (lockres->l_flags & OCFS2_LOCK_BLOCKED &&
1520 !ocfs2_may_continue_on_blocked_lock(lockres, level)) {
1521 /* is the lock is currently blocked on behalf of
1522 * another node */
1523 lockres_add_mask_waiter(lockres, &mw, OCFS2_LOCK_BLOCKED, 0);
1524 wait = 1;
1525 goto unlock;
1526 }
1527
1528 if (level > lockres->l_level) {
1529 if (noqueue_attempted > 0) {
1530 ret = -EAGAIN;
1531 goto unlock;
1532 }
1533 if (lkm_flags & DLM_LKF_NOQUEUE)
1534 noqueue_attempted = 1;
1535
1536 if (lockres->l_action != OCFS2_AST_INVALID)
1537 mlog(ML_ERROR, "lockres %s has action %u pending\n",
1538 lockres->l_name, lockres->l_action);
1539
1540 if (!(lockres->l_flags & OCFS2_LOCK_ATTACHED)) {
1541 lockres->l_action = OCFS2_AST_ATTACH;
1542 lkm_flags &= ~DLM_LKF_CONVERT;
1543 } else {
1544 lockres->l_action = OCFS2_AST_CONVERT;
1545 lkm_flags |= DLM_LKF_CONVERT;
1546 }
1547
1548 lockres->l_requested = level;
1549 lockres_or_flags(lockres, OCFS2_LOCK_BUSY);
1550 gen = lockres_set_pending(lockres);
1551 spin_unlock_irqrestore(&lockres->l_lock, flags);
1552
1553 BUG_ON(level == DLM_LOCK_IV);
1554 BUG_ON(level == DLM_LOCK_NL);
1555
1556 mlog(ML_BASTS, "lockres %s, convert from %d to %d\n",
1557 lockres->l_name, lockres->l_level, level);
1558
1559 /* call dlm_lock to upgrade lock now */
1560 ret = ocfs2_dlm_lock(osb->cconn,
1561 level,
1562 &lockres->l_lksb,
1563 lkm_flags,
1564 lockres->l_name,
1565 OCFS2_LOCK_ID_MAX_LEN - 1);
1566 lockres_clear_pending(lockres, gen, osb);
1567 if (ret) {
1568 if (!(lkm_flags & DLM_LKF_NOQUEUE) ||
1569 (ret != -EAGAIN)) {
1570 ocfs2_log_dlm_error("ocfs2_dlm_lock",
1571 ret, lockres);
1572 }
1573 ocfs2_recover_from_dlm_error(lockres, 1);
1574 goto out;
1575 }
1576 dlm_locked = 1;
1577
1578 mlog(0, "lock %s, successful return from ocfs2_dlm_lock\n",
1579 lockres->l_name);
1580
1581 /* At this point we've gone inside the dlm and need to
1582 * complete our work regardless. */
1583 catch_signals = 0;
1584
1585 /* wait for busy to clear and carry on */
1586 goto again;
1587 }
1588
1589update_holders:
1590 /* Ok, if we get here then we're good to go. */
1591 ocfs2_inc_holders(lockres, level);
1592
1593 ret = 0;
1594unlock:
1595 lockres_clear_flags(lockres, OCFS2_LOCK_UPCONVERT_FINISHING);
1596
1597 /* ocfs2_unblock_lock reques on seeing OCFS2_LOCK_UPCONVERT_FINISHING */
1598 kick_dc = (lockres->l_flags & OCFS2_LOCK_BLOCKED);
1599
1600 spin_unlock_irqrestore(&lockres->l_lock, flags);
1601 if (kick_dc)
1602 ocfs2_wake_downconvert_thread(osb);
1603out:
1604 /*
1605 * This is helping work around a lock inversion between the page lock
1606 * and dlm locks. One path holds the page lock while calling aops
1607 * which block acquiring dlm locks. The voting thread holds dlm
1608 * locks while acquiring page locks while down converting data locks.
1609 * This block is helping an aop path notice the inversion and back
1610 * off to unlock its page lock before trying the dlm lock again.
1611 */
1612 if (wait && arg_flags & OCFS2_LOCK_NONBLOCK &&
1613 mw.mw_mask & (OCFS2_LOCK_BUSY|OCFS2_LOCK_BLOCKED)) {
1614 wait = 0;
1615 spin_lock_irqsave(&lockres->l_lock, flags);
1616 if (__lockres_remove_mask_waiter(lockres, &mw)) {
1617 if (dlm_locked)
1618 lockres_or_flags(lockres,
1619 OCFS2_LOCK_NONBLOCK_FINISHED);
1620 spin_unlock_irqrestore(&lockres->l_lock, flags);
1621 ret = -EAGAIN;
1622 } else {
1623 spin_unlock_irqrestore(&lockres->l_lock, flags);
1624 goto again;
1625 }
1626 }
1627 if (wait) {
1628 ret = ocfs2_wait_for_mask(&mw);
1629 if (ret == 0)
1630 goto again;
1631 mlog_errno(ret);
1632 }
1633 ocfs2_update_lock_stats(lockres, level, &mw, ret);
1634
1635#ifdef CONFIG_DEBUG_LOCK_ALLOC
1636 if (!ret && lockres->l_lockdep_map.key != NULL) {
1637 if (level == DLM_LOCK_PR)
1638 rwsem_acquire_read(&lockres->l_lockdep_map, l_subclass,
1639 !!(arg_flags & OCFS2_META_LOCK_NOQUEUE),
1640 caller_ip);
1641 else
1642 rwsem_acquire(&lockres->l_lockdep_map, l_subclass,
1643 !!(arg_flags & OCFS2_META_LOCK_NOQUEUE),
1644 caller_ip);
1645 }
1646#endif
1647 return ret;
1648}
1649
1650static inline int ocfs2_cluster_lock(struct ocfs2_super *osb,
1651 struct ocfs2_lock_res *lockres,
1652 int level,
1653 u32 lkm_flags,
1654 int arg_flags)
1655{
1656 return __ocfs2_cluster_lock(osb, lockres, level, lkm_flags, arg_flags,
1657 0, _RET_IP_);
1658}
1659
1660
1661static void __ocfs2_cluster_unlock(struct ocfs2_super *osb,
1662 struct ocfs2_lock_res *lockres,
1663 int level,
1664 unsigned long caller_ip)
1665{
1666 unsigned long flags;
1667
1668 spin_lock_irqsave(&lockres->l_lock, flags);
1669 ocfs2_dec_holders(lockres, level);
1670 ocfs2_downconvert_on_unlock(osb, lockres);
1671 spin_unlock_irqrestore(&lockres->l_lock, flags);
1672#ifdef CONFIG_DEBUG_LOCK_ALLOC
1673 if (lockres->l_lockdep_map.key != NULL)
1674 rwsem_release(&lockres->l_lockdep_map, 1, caller_ip);
1675#endif
1676}
1677
1678static int ocfs2_create_new_lock(struct ocfs2_super *osb,
1679 struct ocfs2_lock_res *lockres,
1680 int ex,
1681 int local)
1682{
1683 int level = ex ? DLM_LOCK_EX : DLM_LOCK_PR;
1684 unsigned long flags;
1685 u32 lkm_flags = local ? DLM_LKF_LOCAL : 0;
1686
1687 spin_lock_irqsave(&lockres->l_lock, flags);
1688 BUG_ON(lockres->l_flags & OCFS2_LOCK_ATTACHED);
1689 lockres_or_flags(lockres, OCFS2_LOCK_LOCAL);
1690 spin_unlock_irqrestore(&lockres->l_lock, flags);
1691
1692 return ocfs2_lock_create(osb, lockres, level, lkm_flags);
1693}
1694
1695/* Grants us an EX lock on the data and metadata resources, skipping
1696 * the normal cluster directory lookup. Use this ONLY on newly created
1697 * inodes which other nodes can't possibly see, and which haven't been
1698 * hashed in the inode hash yet. This can give us a good performance
1699 * increase as it'll skip the network broadcast normally associated
1700 * with creating a new lock resource. */
1701int ocfs2_create_new_inode_locks(struct inode *inode)
1702{
1703 int ret;
1704 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1705
1706 BUG_ON(!ocfs2_inode_is_new(inode));
1707
1708 mlog(0, "Inode %llu\n", (unsigned long long)OCFS2_I(inode)->ip_blkno);
1709
1710 /* NOTE: That we don't increment any of the holder counts, nor
1711 * do we add anything to a journal handle. Since this is
1712 * supposed to be a new inode which the cluster doesn't know
1713 * about yet, there is no need to. As far as the LVB handling
1714 * is concerned, this is basically like acquiring an EX lock
1715 * on a resource which has an invalid one -- we'll set it
1716 * valid when we release the EX. */
1717
1718 ret = ocfs2_create_new_lock(osb, &OCFS2_I(inode)->ip_rw_lockres, 1, 1);
1719 if (ret) {
1720 mlog_errno(ret);
1721 goto bail;
1722 }
1723
1724 /*
1725 * We don't want to use DLM_LKF_LOCAL on a meta data lock as they
1726 * don't use a generation in their lock names.
1727 */
1728 ret = ocfs2_create_new_lock(osb, &OCFS2_I(inode)->ip_inode_lockres, 1, 0);
1729 if (ret) {
1730 mlog_errno(ret);
1731 goto bail;
1732 }
1733
1734 ret = ocfs2_create_new_lock(osb, &OCFS2_I(inode)->ip_open_lockres, 0, 0);
1735 if (ret)
1736 mlog_errno(ret);
1737
1738bail:
1739 return ret;
1740}
1741
1742int ocfs2_rw_lock(struct inode *inode, int write)
1743{
1744 int status, level;
1745 struct ocfs2_lock_res *lockres;
1746 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1747
1748 mlog(0, "inode %llu take %s RW lock\n",
1749 (unsigned long long)OCFS2_I(inode)->ip_blkno,
1750 write ? "EXMODE" : "PRMODE");
1751
1752 if (ocfs2_mount_local(osb))
1753 return 0;
1754
1755 lockres = &OCFS2_I(inode)->ip_rw_lockres;
1756
1757 level = write ? DLM_LOCK_EX : DLM_LOCK_PR;
1758
1759 status = ocfs2_cluster_lock(osb, lockres, level, 0, 0);
1760 if (status < 0)
1761 mlog_errno(status);
1762
1763 return status;
1764}
1765
1766int ocfs2_try_rw_lock(struct inode *inode, int write)
1767{
1768 int status, level;
1769 struct ocfs2_lock_res *lockres;
1770 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1771
1772 mlog(0, "inode %llu try to take %s RW lock\n",
1773 (unsigned long long)OCFS2_I(inode)->ip_blkno,
1774 write ? "EXMODE" : "PRMODE");
1775
1776 if (ocfs2_mount_local(osb))
1777 return 0;
1778
1779 lockres = &OCFS2_I(inode)->ip_rw_lockres;
1780
1781 level = write ? DLM_LOCK_EX : DLM_LOCK_PR;
1782
1783 status = ocfs2_cluster_lock(osb, lockres, level, DLM_LKF_NOQUEUE, 0);
1784 return status;
1785}
1786
1787void ocfs2_rw_unlock(struct inode *inode, int write)
1788{
1789 int level = write ? DLM_LOCK_EX : DLM_LOCK_PR;
1790 struct ocfs2_lock_res *lockres = &OCFS2_I(inode)->ip_rw_lockres;
1791 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1792
1793 mlog(0, "inode %llu drop %s RW lock\n",
1794 (unsigned long long)OCFS2_I(inode)->ip_blkno,
1795 write ? "EXMODE" : "PRMODE");
1796
1797 if (!ocfs2_mount_local(osb))
1798 ocfs2_cluster_unlock(osb, lockres, level);
1799}
1800
1801/*
1802 * ocfs2_open_lock always get PR mode lock.
1803 */
1804int ocfs2_open_lock(struct inode *inode)
1805{
1806 int status = 0;
1807 struct ocfs2_lock_res *lockres;
1808 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1809
1810 mlog(0, "inode %llu take PRMODE open lock\n",
1811 (unsigned long long)OCFS2_I(inode)->ip_blkno);
1812
1813 if (ocfs2_is_hard_readonly(osb) || ocfs2_mount_local(osb))
1814 goto out;
1815
1816 lockres = &OCFS2_I(inode)->ip_open_lockres;
1817
1818 status = ocfs2_cluster_lock(osb, lockres, DLM_LOCK_PR, 0, 0);
1819 if (status < 0)
1820 mlog_errno(status);
1821
1822out:
1823 return status;
1824}
1825
1826int ocfs2_try_open_lock(struct inode *inode, int write)
1827{
1828 int status = 0, level;
1829 struct ocfs2_lock_res *lockres;
1830 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1831
1832 mlog(0, "inode %llu try to take %s open lock\n",
1833 (unsigned long long)OCFS2_I(inode)->ip_blkno,
1834 write ? "EXMODE" : "PRMODE");
1835
1836 if (ocfs2_is_hard_readonly(osb)) {
1837 if (write)
1838 status = -EROFS;
1839 goto out;
1840 }
1841
1842 if (ocfs2_mount_local(osb))
1843 goto out;
1844
1845 lockres = &OCFS2_I(inode)->ip_open_lockres;
1846
1847 level = write ? DLM_LOCK_EX : DLM_LOCK_PR;
1848
1849 /*
1850 * The file system may already holding a PRMODE/EXMODE open lock.
1851 * Since we pass DLM_LKF_NOQUEUE, the request won't block waiting on
1852 * other nodes and the -EAGAIN will indicate to the caller that
1853 * this inode is still in use.
1854 */
1855 status = ocfs2_cluster_lock(osb, lockres, level, DLM_LKF_NOQUEUE, 0);
1856
1857out:
1858 return status;
1859}
1860
1861/*
1862 * ocfs2_open_unlock unlock PR and EX mode open locks.
1863 */
1864void ocfs2_open_unlock(struct inode *inode)
1865{
1866 struct ocfs2_lock_res *lockres = &OCFS2_I(inode)->ip_open_lockres;
1867 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1868
1869 mlog(0, "inode %llu drop open lock\n",
1870 (unsigned long long)OCFS2_I(inode)->ip_blkno);
1871
1872 if (ocfs2_mount_local(osb))
1873 goto out;
1874
1875 if(lockres->l_ro_holders)
1876 ocfs2_cluster_unlock(osb, lockres, DLM_LOCK_PR);
1877 if(lockres->l_ex_holders)
1878 ocfs2_cluster_unlock(osb, lockres, DLM_LOCK_EX);
1879
1880out:
1881 return;
1882}
1883
1884static int ocfs2_flock_handle_signal(struct ocfs2_lock_res *lockres,
1885 int level)
1886{
1887 int ret;
1888 struct ocfs2_super *osb = ocfs2_get_lockres_osb(lockres);
1889 unsigned long flags;
1890 struct ocfs2_mask_waiter mw;
1891
1892 ocfs2_init_mask_waiter(&mw);
1893
1894retry_cancel:
1895 spin_lock_irqsave(&lockres->l_lock, flags);
1896 if (lockres->l_flags & OCFS2_LOCK_BUSY) {
1897 ret = ocfs2_prepare_cancel_convert(osb, lockres);
1898 if (ret) {
1899 spin_unlock_irqrestore(&lockres->l_lock, flags);
1900 ret = ocfs2_cancel_convert(osb, lockres);
1901 if (ret < 0) {
1902 mlog_errno(ret);
1903 goto out;
1904 }
1905 goto retry_cancel;
1906 }
1907 lockres_add_mask_waiter(lockres, &mw, OCFS2_LOCK_BUSY, 0);
1908 spin_unlock_irqrestore(&lockres->l_lock, flags);
1909
1910 ocfs2_wait_for_mask(&mw);
1911 goto retry_cancel;
1912 }
1913
1914 ret = -ERESTARTSYS;
1915 /*
1916 * We may still have gotten the lock, in which case there's no
1917 * point to restarting the syscall.
1918 */
1919 if (lockres->l_level == level)
1920 ret = 0;
1921
1922 mlog(0, "Cancel returning %d. flags: 0x%lx, level: %d, act: %d\n", ret,
1923 lockres->l_flags, lockres->l_level, lockres->l_action);
1924
1925 spin_unlock_irqrestore(&lockres->l_lock, flags);
1926
1927out:
1928 return ret;
1929}
1930
1931/*
1932 * ocfs2_file_lock() and ocfs2_file_unlock() map to a single pair of
1933 * flock() calls. The locking approach this requires is sufficiently
1934 * different from all other cluster lock types that we implement a
1935 * separate path to the "low-level" dlm calls. In particular:
1936 *
1937 * - No optimization of lock levels is done - we take at exactly
1938 * what's been requested.
1939 *
1940 * - No lock caching is employed. We immediately downconvert to
1941 * no-lock at unlock time. This also means flock locks never go on
1942 * the blocking list).
1943 *
1944 * - Since userspace can trivially deadlock itself with flock, we make
1945 * sure to allow cancellation of a misbehaving applications flock()
1946 * request.
1947 *
1948 * - Access to any flock lockres doesn't require concurrency, so we
1949 * can simplify the code by requiring the caller to guarantee
1950 * serialization of dlmglue flock calls.
1951 */
1952int ocfs2_file_lock(struct file *file, int ex, int trylock)
1953{
1954 int ret, level = ex ? DLM_LOCK_EX : DLM_LOCK_PR;
1955 unsigned int lkm_flags = trylock ? DLM_LKF_NOQUEUE : 0;
1956 unsigned long flags;
1957 struct ocfs2_file_private *fp = file->private_data;
1958 struct ocfs2_lock_res *lockres = &fp->fp_flock;
1959 struct ocfs2_super *osb = OCFS2_SB(file->f_mapping->host->i_sb);
1960 struct ocfs2_mask_waiter mw;
1961
1962 ocfs2_init_mask_waiter(&mw);
1963
1964 if ((lockres->l_flags & OCFS2_LOCK_BUSY) ||
1965 (lockres->l_level > DLM_LOCK_NL)) {
1966 mlog(ML_ERROR,
1967 "File lock \"%s\" has busy or locked state: flags: 0x%lx, "
1968 "level: %u\n", lockres->l_name, lockres->l_flags,
1969 lockres->l_level);
1970 return -EINVAL;
1971 }
1972
1973 spin_lock_irqsave(&lockres->l_lock, flags);
1974 if (!(lockres->l_flags & OCFS2_LOCK_ATTACHED)) {
1975 lockres_add_mask_waiter(lockres, &mw, OCFS2_LOCK_BUSY, 0);
1976 spin_unlock_irqrestore(&lockres->l_lock, flags);
1977
1978 /*
1979 * Get the lock at NLMODE to start - that way we
1980 * can cancel the upconvert request if need be.
1981 */
1982 ret = ocfs2_lock_create(osb, lockres, DLM_LOCK_NL, 0);
1983 if (ret < 0) {
1984 mlog_errno(ret);
1985 goto out;
1986 }
1987
1988 ret = ocfs2_wait_for_mask(&mw);
1989 if (ret) {
1990 mlog_errno(ret);
1991 goto out;
1992 }
1993 spin_lock_irqsave(&lockres->l_lock, flags);
1994 }
1995
1996 lockres->l_action = OCFS2_AST_CONVERT;
1997 lkm_flags |= DLM_LKF_CONVERT;
1998 lockres->l_requested = level;
1999 lockres_or_flags(lockres, OCFS2_LOCK_BUSY);
2000
2001 lockres_add_mask_waiter(lockres, &mw, OCFS2_LOCK_BUSY, 0);
2002 spin_unlock_irqrestore(&lockres->l_lock, flags);
2003
2004 ret = ocfs2_dlm_lock(osb->cconn, level, &lockres->l_lksb, lkm_flags,
2005 lockres->l_name, OCFS2_LOCK_ID_MAX_LEN - 1);
2006 if (ret) {
2007 if (!trylock || (ret != -EAGAIN)) {
2008 ocfs2_log_dlm_error("ocfs2_dlm_lock", ret, lockres);
2009 ret = -EINVAL;
2010 }
2011
2012 ocfs2_recover_from_dlm_error(lockres, 1);
2013 lockres_remove_mask_waiter(lockres, &mw);
2014 goto out;
2015 }
2016
2017 ret = ocfs2_wait_for_mask_interruptible(&mw, lockres);
2018 if (ret == -ERESTARTSYS) {
2019 /*
2020 * Userspace can cause deadlock itself with
2021 * flock(). Current behavior locally is to allow the
2022 * deadlock, but abort the system call if a signal is
2023 * received. We follow this example, otherwise a
2024 * poorly written program could sit in kernel until
2025 * reboot.
2026 *
2027 * Handling this is a bit more complicated for Ocfs2
2028 * though. We can't exit this function with an
2029 * outstanding lock request, so a cancel convert is
2030 * required. We intentionally overwrite 'ret' - if the
2031 * cancel fails and the lock was granted, it's easier
2032 * to just bubble success back up to the user.
2033 */
2034 ret = ocfs2_flock_handle_signal(lockres, level);
2035 } else if (!ret && (level > lockres->l_level)) {
2036 /* Trylock failed asynchronously */
2037 BUG_ON(!trylock);
2038 ret = -EAGAIN;
2039 }
2040
2041out:
2042
2043 mlog(0, "Lock: \"%s\" ex: %d, trylock: %d, returns: %d\n",
2044 lockres->l_name, ex, trylock, ret);
2045 return ret;
2046}
2047
2048void ocfs2_file_unlock(struct file *file)
2049{
2050 int ret;
2051 unsigned int gen;
2052 unsigned long flags;
2053 struct ocfs2_file_private *fp = file->private_data;
2054 struct ocfs2_lock_res *lockres = &fp->fp_flock;
2055 struct ocfs2_super *osb = OCFS2_SB(file->f_mapping->host->i_sb);
2056 struct ocfs2_mask_waiter mw;
2057
2058 ocfs2_init_mask_waiter(&mw);
2059
2060 if (!(lockres->l_flags & OCFS2_LOCK_ATTACHED))
2061 return;
2062
2063 if (lockres->l_level == DLM_LOCK_NL)
2064 return;
2065
2066 mlog(0, "Unlock: \"%s\" flags: 0x%lx, level: %d, act: %d\n",
2067 lockres->l_name, lockres->l_flags, lockres->l_level,
2068 lockres->l_action);
2069
2070 spin_lock_irqsave(&lockres->l_lock, flags);
2071 /*
2072 * Fake a blocking ast for the downconvert code.
2073 */
2074 lockres_or_flags(lockres, OCFS2_LOCK_BLOCKED);
2075 lockres->l_blocking = DLM_LOCK_EX;
2076
2077 gen = ocfs2_prepare_downconvert(lockres, DLM_LOCK_NL);
2078 lockres_add_mask_waiter(lockres, &mw, OCFS2_LOCK_BUSY, 0);
2079 spin_unlock_irqrestore(&lockres->l_lock, flags);
2080
2081 ret = ocfs2_downconvert_lock(osb, lockres, DLM_LOCK_NL, 0, gen);
2082 if (ret) {
2083 mlog_errno(ret);
2084 return;
2085 }
2086
2087 ret = ocfs2_wait_for_mask(&mw);
2088 if (ret)
2089 mlog_errno(ret);
2090}
2091
2092static void ocfs2_downconvert_on_unlock(struct ocfs2_super *osb,
2093 struct ocfs2_lock_res *lockres)
2094{
2095 int kick = 0;
2096
2097 /* If we know that another node is waiting on our lock, kick
2098 * the downconvert thread * pre-emptively when we reach a release
2099 * condition. */
2100 if (lockres->l_flags & OCFS2_LOCK_BLOCKED) {
2101 switch(lockres->l_blocking) {
2102 case DLM_LOCK_EX:
2103 if (!lockres->l_ex_holders && !lockres->l_ro_holders)
2104 kick = 1;
2105 break;
2106 case DLM_LOCK_PR:
2107 if (!lockres->l_ex_holders)
2108 kick = 1;
2109 break;
2110 default:
2111 BUG();
2112 }
2113 }
2114
2115 if (kick)
2116 ocfs2_wake_downconvert_thread(osb);
2117}
2118
2119#define OCFS2_SEC_BITS 34
2120#define OCFS2_SEC_SHIFT (64 - 34)
2121#define OCFS2_NSEC_MASK ((1ULL << OCFS2_SEC_SHIFT) - 1)
2122
2123/* LVB only has room for 64 bits of time here so we pack it for
2124 * now. */
2125static u64 ocfs2_pack_timespec(struct timespec *spec)
2126{
2127 u64 res;
2128 u64 sec = spec->tv_sec;
2129 u32 nsec = spec->tv_nsec;
2130
2131 res = (sec << OCFS2_SEC_SHIFT) | (nsec & OCFS2_NSEC_MASK);
2132
2133 return res;
2134}
2135
2136/* Call this with the lockres locked. I am reasonably sure we don't
2137 * need ip_lock in this function as anyone who would be changing those
2138 * values is supposed to be blocked in ocfs2_inode_lock right now. */
2139static void __ocfs2_stuff_meta_lvb(struct inode *inode)
2140{
2141 struct ocfs2_inode_info *oi = OCFS2_I(inode);
2142 struct ocfs2_lock_res *lockres = &oi->ip_inode_lockres;
2143 struct ocfs2_meta_lvb *lvb;
2144
2145 lvb = ocfs2_dlm_lvb(&lockres->l_lksb);
2146
2147 /*
2148 * Invalidate the LVB of a deleted inode - this way other
2149 * nodes are forced to go to disk and discover the new inode
2150 * status.
2151 */
2152 if (oi->ip_flags & OCFS2_INODE_DELETED) {
2153 lvb->lvb_version = 0;
2154 goto out;
2155 }
2156
2157 lvb->lvb_version = OCFS2_LVB_VERSION;
2158 lvb->lvb_isize = cpu_to_be64(i_size_read(inode));
2159 lvb->lvb_iclusters = cpu_to_be32(oi->ip_clusters);
2160 lvb->lvb_iuid = cpu_to_be32(i_uid_read(inode));
2161 lvb->lvb_igid = cpu_to_be32(i_gid_read(inode));
2162 lvb->lvb_imode = cpu_to_be16(inode->i_mode);
2163 lvb->lvb_inlink = cpu_to_be16(inode->i_nlink);
2164 lvb->lvb_iatime_packed =
2165 cpu_to_be64(ocfs2_pack_timespec(&inode->i_atime));
2166 lvb->lvb_ictime_packed =
2167 cpu_to_be64(ocfs2_pack_timespec(&inode->i_ctime));
2168 lvb->lvb_imtime_packed =
2169 cpu_to_be64(ocfs2_pack_timespec(&inode->i_mtime));
2170 lvb->lvb_iattr = cpu_to_be32(oi->ip_attr);
2171 lvb->lvb_idynfeatures = cpu_to_be16(oi->ip_dyn_features);
2172 lvb->lvb_igeneration = cpu_to_be32(inode->i_generation);
2173
2174out:
2175 mlog_meta_lvb(0, lockres);
2176}
2177
2178static void ocfs2_unpack_timespec(struct timespec *spec,
2179 u64 packed_time)
2180{
2181 spec->tv_sec = packed_time >> OCFS2_SEC_SHIFT;
2182 spec->tv_nsec = packed_time & OCFS2_NSEC_MASK;
2183}
2184
2185static void ocfs2_refresh_inode_from_lvb(struct inode *inode)
2186{
2187 struct ocfs2_inode_info *oi = OCFS2_I(inode);
2188 struct ocfs2_lock_res *lockres = &oi->ip_inode_lockres;
2189 struct ocfs2_meta_lvb *lvb;
2190
2191 mlog_meta_lvb(0, lockres);
2192
2193 lvb = ocfs2_dlm_lvb(&lockres->l_lksb);
2194
2195 /* We're safe here without the lockres lock... */
2196 spin_lock(&oi->ip_lock);
2197 oi->ip_clusters = be32_to_cpu(lvb->lvb_iclusters);
2198 i_size_write(inode, be64_to_cpu(lvb->lvb_isize));
2199
2200 oi->ip_attr = be32_to_cpu(lvb->lvb_iattr);
2201 oi->ip_dyn_features = be16_to_cpu(lvb->lvb_idynfeatures);
2202 ocfs2_set_inode_flags(inode);
2203
2204 /* fast-symlinks are a special case */
2205 if (S_ISLNK(inode->i_mode) && !oi->ip_clusters)
2206 inode->i_blocks = 0;
2207 else
2208 inode->i_blocks = ocfs2_inode_sector_count(inode);
2209
2210 i_uid_write(inode, be32_to_cpu(lvb->lvb_iuid));
2211 i_gid_write(inode, be32_to_cpu(lvb->lvb_igid));
2212 inode->i_mode = be16_to_cpu(lvb->lvb_imode);
2213 set_nlink(inode, be16_to_cpu(lvb->lvb_inlink));
2214 ocfs2_unpack_timespec(&inode->i_atime,
2215 be64_to_cpu(lvb->lvb_iatime_packed));
2216 ocfs2_unpack_timespec(&inode->i_mtime,
2217 be64_to_cpu(lvb->lvb_imtime_packed));
2218 ocfs2_unpack_timespec(&inode->i_ctime,
2219 be64_to_cpu(lvb->lvb_ictime_packed));
2220 spin_unlock(&oi->ip_lock);
2221}
2222
2223static inline int ocfs2_meta_lvb_is_trustable(struct inode *inode,
2224 struct ocfs2_lock_res *lockres)
2225{
2226 struct ocfs2_meta_lvb *lvb = ocfs2_dlm_lvb(&lockres->l_lksb);
2227
2228 if (ocfs2_dlm_lvb_valid(&lockres->l_lksb)
2229 && lvb->lvb_version == OCFS2_LVB_VERSION
2230 && be32_to_cpu(lvb->lvb_igeneration) == inode->i_generation)
2231 return 1;
2232 return 0;
2233}
2234
2235/* Determine whether a lock resource needs to be refreshed, and
2236 * arbitrate who gets to refresh it.
2237 *
2238 * 0 means no refresh needed.
2239 *
2240 * > 0 means you need to refresh this and you MUST call
2241 * ocfs2_complete_lock_res_refresh afterwards. */
2242static int ocfs2_should_refresh_lock_res(struct ocfs2_lock_res *lockres)
2243{
2244 unsigned long flags;
2245 int status = 0;
2246
2247refresh_check:
2248 spin_lock_irqsave(&lockres->l_lock, flags);
2249 if (!(lockres->l_flags & OCFS2_LOCK_NEEDS_REFRESH)) {
2250 spin_unlock_irqrestore(&lockres->l_lock, flags);
2251 goto bail;
2252 }
2253
2254 if (lockres->l_flags & OCFS2_LOCK_REFRESHING) {
2255 spin_unlock_irqrestore(&lockres->l_lock, flags);
2256
2257 ocfs2_wait_on_refreshing_lock(lockres);
2258 goto refresh_check;
2259 }
2260
2261 /* Ok, I'll be the one to refresh this lock. */
2262 lockres_or_flags(lockres, OCFS2_LOCK_REFRESHING);
2263 spin_unlock_irqrestore(&lockres->l_lock, flags);
2264
2265 status = 1;
2266bail:
2267 mlog(0, "status %d\n", status);
2268 return status;
2269}
2270
2271/* If status is non zero, I'll mark it as not being in refresh
2272 * anymroe, but i won't clear the needs refresh flag. */
2273static inline void ocfs2_complete_lock_res_refresh(struct ocfs2_lock_res *lockres,
2274 int status)
2275{
2276 unsigned long flags;
2277
2278 spin_lock_irqsave(&lockres->l_lock, flags);
2279 lockres_clear_flags(lockres, OCFS2_LOCK_REFRESHING);
2280 if (!status)
2281 lockres_clear_flags(lockres, OCFS2_LOCK_NEEDS_REFRESH);
2282 spin_unlock_irqrestore(&lockres->l_lock, flags);
2283
2284 wake_up(&lockres->l_event);
2285}
2286
2287/* may or may not return a bh if it went to disk. */
2288static int ocfs2_inode_lock_update(struct inode *inode,
2289 struct buffer_head **bh)
2290{
2291 int status = 0;
2292 struct ocfs2_inode_info *oi = OCFS2_I(inode);
2293 struct ocfs2_lock_res *lockres = &oi->ip_inode_lockres;
2294 struct ocfs2_dinode *fe;
2295 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
2296
2297 if (ocfs2_mount_local(osb))
2298 goto bail;
2299
2300 spin_lock(&oi->ip_lock);
2301 if (oi->ip_flags & OCFS2_INODE_DELETED) {
2302 mlog(0, "Orphaned inode %llu was deleted while we "
2303 "were waiting on a lock. ip_flags = 0x%x\n",
2304 (unsigned long long)oi->ip_blkno, oi->ip_flags);
2305 spin_unlock(&oi->ip_lock);
2306 status = -ENOENT;
2307 goto bail;
2308 }
2309 spin_unlock(&oi->ip_lock);
2310
2311 if (!ocfs2_should_refresh_lock_res(lockres))
2312 goto bail;
2313
2314 /* This will discard any caching information we might have had
2315 * for the inode metadata. */
2316 ocfs2_metadata_cache_purge(INODE_CACHE(inode));
2317
2318 ocfs2_extent_map_trunc(inode, 0);
2319
2320 if (ocfs2_meta_lvb_is_trustable(inode, lockres)) {
2321 mlog(0, "Trusting LVB on inode %llu\n",
2322 (unsigned long long)oi->ip_blkno);
2323 ocfs2_refresh_inode_from_lvb(inode);
2324 } else {
2325 /* Boo, we have to go to disk. */
2326 /* read bh, cast, ocfs2_refresh_inode */
2327 status = ocfs2_read_inode_block(inode, bh);
2328 if (status < 0) {
2329 mlog_errno(status);
2330 goto bail_refresh;
2331 }
2332 fe = (struct ocfs2_dinode *) (*bh)->b_data;
2333
2334 /* This is a good chance to make sure we're not
2335 * locking an invalid object. ocfs2_read_inode_block()
2336 * already checked that the inode block is sane.
2337 *
2338 * We bug on a stale inode here because we checked
2339 * above whether it was wiped from disk. The wiping
2340 * node provides a guarantee that we receive that
2341 * message and can mark the inode before dropping any
2342 * locks associated with it. */
2343 mlog_bug_on_msg(inode->i_generation !=
2344 le32_to_cpu(fe->i_generation),
2345 "Invalid dinode %llu disk generation: %u "
2346 "inode->i_generation: %u\n",
2347 (unsigned long long)oi->ip_blkno,
2348 le32_to_cpu(fe->i_generation),
2349 inode->i_generation);
2350 mlog_bug_on_msg(le64_to_cpu(fe->i_dtime) ||
2351 !(fe->i_flags & cpu_to_le32(OCFS2_VALID_FL)),
2352 "Stale dinode %llu dtime: %llu flags: 0x%x\n",
2353 (unsigned long long)oi->ip_blkno,
2354 (unsigned long long)le64_to_cpu(fe->i_dtime),
2355 le32_to_cpu(fe->i_flags));
2356
2357 ocfs2_refresh_inode(inode, fe);
2358 ocfs2_track_lock_refresh(lockres);
2359 }
2360
2361 status = 0;
2362bail_refresh:
2363 ocfs2_complete_lock_res_refresh(lockres, status);
2364bail:
2365 return status;
2366}
2367
2368static int ocfs2_assign_bh(struct inode *inode,
2369 struct buffer_head **ret_bh,
2370 struct buffer_head *passed_bh)
2371{
2372 int status;
2373
2374 if (passed_bh) {
2375 /* Ok, the update went to disk for us, use the
2376 * returned bh. */
2377 *ret_bh = passed_bh;
2378 get_bh(*ret_bh);
2379
2380 return 0;
2381 }
2382
2383 status = ocfs2_read_inode_block(inode, ret_bh);
2384 if (status < 0)
2385 mlog_errno(status);
2386
2387 return status;
2388}
2389
2390/*
2391 * returns < 0 error if the callback will never be called, otherwise
2392 * the result of the lock will be communicated via the callback.
2393 */
2394int ocfs2_inode_lock_full_nested(struct inode *inode,
2395 struct buffer_head **ret_bh,
2396 int ex,
2397 int arg_flags,
2398 int subclass)
2399{
2400 int status, level, acquired;
2401 u32 dlm_flags;
2402 struct ocfs2_lock_res *lockres = NULL;
2403 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
2404 struct buffer_head *local_bh = NULL;
2405
2406 mlog(0, "inode %llu, take %s META lock\n",
2407 (unsigned long long)OCFS2_I(inode)->ip_blkno,
2408 ex ? "EXMODE" : "PRMODE");
2409
2410 status = 0;
2411 acquired = 0;
2412 /* We'll allow faking a readonly metadata lock for
2413 * rodevices. */
2414 if (ocfs2_is_hard_readonly(osb)) {
2415 if (ex)
2416 status = -EROFS;
2417 goto getbh;
2418 }
2419
2420 if ((arg_flags & OCFS2_META_LOCK_GETBH) ||
2421 ocfs2_mount_local(osb))
2422 goto update;
2423
2424 if (!(arg_flags & OCFS2_META_LOCK_RECOVERY))
2425 ocfs2_wait_for_recovery(osb);
2426
2427 lockres = &OCFS2_I(inode)->ip_inode_lockres;
2428 level = ex ? DLM_LOCK_EX : DLM_LOCK_PR;
2429 dlm_flags = 0;
2430 if (arg_flags & OCFS2_META_LOCK_NOQUEUE)
2431 dlm_flags |= DLM_LKF_NOQUEUE;
2432
2433 status = __ocfs2_cluster_lock(osb, lockres, level, dlm_flags,
2434 arg_flags, subclass, _RET_IP_);
2435 if (status < 0) {
2436 if (status != -EAGAIN)
2437 mlog_errno(status);
2438 goto bail;
2439 }
2440
2441 /* Notify the error cleanup path to drop the cluster lock. */
2442 acquired = 1;
2443
2444 /* We wait twice because a node may have died while we were in
2445 * the lower dlm layers. The second time though, we've
2446 * committed to owning this lock so we don't allow signals to
2447 * abort the operation. */
2448 if (!(arg_flags & OCFS2_META_LOCK_RECOVERY))
2449 ocfs2_wait_for_recovery(osb);
2450
2451update:
2452 /*
2453 * We only see this flag if we're being called from
2454 * ocfs2_read_locked_inode(). It means we're locking an inode
2455 * which hasn't been populated yet, so clear the refresh flag
2456 * and let the caller handle it.
2457 */
2458 if (inode->i_state & I_NEW) {
2459 status = 0;
2460 if (lockres)
2461 ocfs2_complete_lock_res_refresh(lockres, 0);
2462 goto bail;
2463 }
2464
2465 /* This is fun. The caller may want a bh back, or it may
2466 * not. ocfs2_inode_lock_update definitely wants one in, but
2467 * may or may not read one, depending on what's in the
2468 * LVB. The result of all of this is that we've *only* gone to
2469 * disk if we have to, so the complexity is worthwhile. */
2470 status = ocfs2_inode_lock_update(inode, &local_bh);
2471 if (status < 0) {
2472 if (status != -ENOENT)
2473 mlog_errno(status);
2474 goto bail;
2475 }
2476getbh:
2477 if (ret_bh) {
2478 status = ocfs2_assign_bh(inode, ret_bh, local_bh);
2479 if (status < 0) {
2480 mlog_errno(status);
2481 goto bail;
2482 }
2483 }
2484
2485bail:
2486 if (status < 0) {
2487 if (ret_bh && (*ret_bh)) {
2488 brelse(*ret_bh);
2489 *ret_bh = NULL;
2490 }
2491 if (acquired)
2492 ocfs2_inode_unlock(inode, ex);
2493 }
2494
2495 if (local_bh)
2496 brelse(local_bh);
2497
2498 return status;
2499}
2500
2501/*
2502 * This is working around a lock inversion between tasks acquiring DLM
2503 * locks while holding a page lock and the downconvert thread which
2504 * blocks dlm lock acquiry while acquiring page locks.
2505 *
2506 * ** These _with_page variantes are only intended to be called from aop
2507 * methods that hold page locks and return a very specific *positive* error
2508 * code that aop methods pass up to the VFS -- test for errors with != 0. **
2509 *
2510 * The DLM is called such that it returns -EAGAIN if it would have
2511 * blocked waiting for the downconvert thread. In that case we unlock
2512 * our page so the downconvert thread can make progress. Once we've
2513 * done this we have to return AOP_TRUNCATED_PAGE so the aop method
2514 * that called us can bubble that back up into the VFS who will then
2515 * immediately retry the aop call.
2516 */
2517int ocfs2_inode_lock_with_page(struct inode *inode,
2518 struct buffer_head **ret_bh,
2519 int ex,
2520 struct page *page)
2521{
2522 int ret;
2523
2524 ret = ocfs2_inode_lock_full(inode, ret_bh, ex, OCFS2_LOCK_NONBLOCK);
2525 if (ret == -EAGAIN) {
2526 unlock_page(page);
2527 /*
2528 * If we can't get inode lock immediately, we should not return
2529 * directly here, since this will lead to a softlockup problem.
2530 * The method is to get a blocking lock and immediately unlock
2531 * before returning, this can avoid CPU resource waste due to
2532 * lots of retries, and benefits fairness in getting lock.
2533 */
2534 if (ocfs2_inode_lock(inode, ret_bh, ex) == 0)
2535 ocfs2_inode_unlock(inode, ex);
2536 ret = AOP_TRUNCATED_PAGE;
2537 }
2538
2539 return ret;
2540}
2541
2542int ocfs2_inode_lock_atime(struct inode *inode,
2543 struct vfsmount *vfsmnt,
2544 int *level, int wait)
2545{
2546 int ret;
2547
2548 if (wait)
2549 ret = ocfs2_inode_lock(inode, NULL, 0);
2550 else
2551 ret = ocfs2_try_inode_lock(inode, NULL, 0);
2552
2553 if (ret < 0) {
2554 if (ret != -EAGAIN)
2555 mlog_errno(ret);
2556 return ret;
2557 }
2558
2559 /*
2560 * If we should update atime, we will get EX lock,
2561 * otherwise we just get PR lock.
2562 */
2563 if (ocfs2_should_update_atime(inode, vfsmnt)) {
2564 struct buffer_head *bh = NULL;
2565
2566 ocfs2_inode_unlock(inode, 0);
2567 if (wait)
2568 ret = ocfs2_inode_lock(inode, &bh, 1);
2569 else
2570 ret = ocfs2_try_inode_lock(inode, &bh, 1);
2571
2572 if (ret < 0) {
2573 if (ret != -EAGAIN)
2574 mlog_errno(ret);
2575 return ret;
2576 }
2577 *level = 1;
2578 if (ocfs2_should_update_atime(inode, vfsmnt))
2579 ocfs2_update_inode_atime(inode, bh);
2580 if (bh)
2581 brelse(bh);
2582 } else
2583 *level = 0;
2584
2585 return ret;
2586}
2587
2588void ocfs2_inode_unlock(struct inode *inode,
2589 int ex)
2590{
2591 int level = ex ? DLM_LOCK_EX : DLM_LOCK_PR;
2592 struct ocfs2_lock_res *lockres = &OCFS2_I(inode)->ip_inode_lockres;
2593 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
2594
2595 mlog(0, "inode %llu drop %s META lock\n",
2596 (unsigned long long)OCFS2_I(inode)->ip_blkno,
2597 ex ? "EXMODE" : "PRMODE");
2598
2599 if (!ocfs2_is_hard_readonly(osb) &&
2600 !ocfs2_mount_local(osb))
2601 ocfs2_cluster_unlock(osb, lockres, level);
2602}
2603
2604/*
2605 * This _tracker variantes are introduced to deal with the recursive cluster
2606 * locking issue. The idea is to keep track of a lock holder on the stack of
2607 * the current process. If there's a lock holder on the stack, we know the
2608 * task context is already protected by cluster locking. Currently, they're
2609 * used in some VFS entry routines.
2610 *
2611 * return < 0 on error, return == 0 if there's no lock holder on the stack
2612 * before this call, return == 1 if this call would be a recursive locking.
2613 */
2614int ocfs2_inode_lock_tracker(struct inode *inode,
2615 struct buffer_head **ret_bh,
2616 int ex,
2617 struct ocfs2_lock_holder *oh)
2618{
2619 int status;
2620 int arg_flags = 0, has_locked;
2621 struct ocfs2_lock_res *lockres;
2622
2623 lockres = &OCFS2_I(inode)->ip_inode_lockres;
2624 has_locked = ocfs2_is_locked_by_me(lockres);
2625 /* Just get buffer head if the cluster lock has been taken */
2626 if (has_locked)
2627 arg_flags = OCFS2_META_LOCK_GETBH;
2628
2629 if (likely(!has_locked || ret_bh)) {
2630 status = ocfs2_inode_lock_full(inode, ret_bh, ex, arg_flags);
2631 if (status < 0) {
2632 if (status != -ENOENT)
2633 mlog_errno(status);
2634 return status;
2635 }
2636 }
2637 if (!has_locked)
2638 ocfs2_add_holder(lockres, oh);
2639
2640 return has_locked;
2641}
2642
2643void ocfs2_inode_unlock_tracker(struct inode *inode,
2644 int ex,
2645 struct ocfs2_lock_holder *oh,
2646 int had_lock)
2647{
2648 struct ocfs2_lock_res *lockres;
2649
2650 lockres = &OCFS2_I(inode)->ip_inode_lockres;
2651 /* had_lock means that the currect process already takes the cluster
2652 * lock previously. If had_lock is 1, we have nothing to do here, and
2653 * it will get unlocked where we got the lock.
2654 */
2655 if (!had_lock) {
2656 ocfs2_remove_holder(lockres, oh);
2657 ocfs2_inode_unlock(inode, ex);
2658 }
2659}
2660
2661int ocfs2_orphan_scan_lock(struct ocfs2_super *osb, u32 *seqno)
2662{
2663 struct ocfs2_lock_res *lockres;
2664 struct ocfs2_orphan_scan_lvb *lvb;
2665 int status = 0;
2666
2667 if (ocfs2_is_hard_readonly(osb))
2668 return -EROFS;
2669
2670 if (ocfs2_mount_local(osb))
2671 return 0;
2672
2673 lockres = &osb->osb_orphan_scan.os_lockres;
2674 status = ocfs2_cluster_lock(osb, lockres, DLM_LOCK_EX, 0, 0);
2675 if (status < 0)
2676 return status;
2677
2678 lvb = ocfs2_dlm_lvb(&lockres->l_lksb);
2679 if (ocfs2_dlm_lvb_valid(&lockres->l_lksb) &&
2680 lvb->lvb_version == OCFS2_ORPHAN_LVB_VERSION)
2681 *seqno = be32_to_cpu(lvb->lvb_os_seqno);
2682 else
2683 *seqno = osb->osb_orphan_scan.os_seqno + 1;
2684
2685 return status;
2686}
2687
2688void ocfs2_orphan_scan_unlock(struct ocfs2_super *osb, u32 seqno)
2689{
2690 struct ocfs2_lock_res *lockres;
2691 struct ocfs2_orphan_scan_lvb *lvb;
2692
2693 if (!ocfs2_is_hard_readonly(osb) && !ocfs2_mount_local(osb)) {
2694 lockres = &osb->osb_orphan_scan.os_lockres;
2695 lvb = ocfs2_dlm_lvb(&lockres->l_lksb);
2696 lvb->lvb_version = OCFS2_ORPHAN_LVB_VERSION;
2697 lvb->lvb_os_seqno = cpu_to_be32(seqno);
2698 ocfs2_cluster_unlock(osb, lockres, DLM_LOCK_EX);
2699 }
2700}
2701
2702int ocfs2_super_lock(struct ocfs2_super *osb,
2703 int ex)
2704{
2705 int status = 0;
2706 int level = ex ? DLM_LOCK_EX : DLM_LOCK_PR;
2707 struct ocfs2_lock_res *lockres = &osb->osb_super_lockres;
2708
2709 if (ocfs2_is_hard_readonly(osb))
2710 return -EROFS;
2711
2712 if (ocfs2_mount_local(osb))
2713 goto bail;
2714
2715 status = ocfs2_cluster_lock(osb, lockres, level, 0, 0);
2716 if (status < 0) {
2717 mlog_errno(status);
2718 goto bail;
2719 }
2720
2721 /* The super block lock path is really in the best position to
2722 * know when resources covered by the lock need to be
2723 * refreshed, so we do it here. Of course, making sense of
2724 * everything is up to the caller :) */
2725 status = ocfs2_should_refresh_lock_res(lockres);
2726 if (status) {
2727 status = ocfs2_refresh_slot_info(osb);
2728
2729 ocfs2_complete_lock_res_refresh(lockres, status);
2730
2731 if (status < 0) {
2732 ocfs2_cluster_unlock(osb, lockres, level);
2733 mlog_errno(status);
2734 }
2735 ocfs2_track_lock_refresh(lockres);
2736 }
2737bail:
2738 return status;
2739}
2740
2741void ocfs2_super_unlock(struct ocfs2_super *osb,
2742 int ex)
2743{
2744 int level = ex ? DLM_LOCK_EX : DLM_LOCK_PR;
2745 struct ocfs2_lock_res *lockres = &osb->osb_super_lockres;
2746
2747 if (!ocfs2_mount_local(osb))
2748 ocfs2_cluster_unlock(osb, lockres, level);
2749}
2750
2751int ocfs2_rename_lock(struct ocfs2_super *osb)
2752{
2753 int status;
2754 struct ocfs2_lock_res *lockres = &osb->osb_rename_lockres;
2755
2756 if (ocfs2_is_hard_readonly(osb))
2757 return -EROFS;
2758
2759 if (ocfs2_mount_local(osb))
2760 return 0;
2761
2762 status = ocfs2_cluster_lock(osb, lockres, DLM_LOCK_EX, 0, 0);
2763 if (status < 0)
2764 mlog_errno(status);
2765
2766 return status;
2767}
2768
2769void ocfs2_rename_unlock(struct ocfs2_super *osb)
2770{
2771 struct ocfs2_lock_res *lockres = &osb->osb_rename_lockres;
2772
2773 if (!ocfs2_mount_local(osb))
2774 ocfs2_cluster_unlock(osb, lockres, DLM_LOCK_EX);
2775}
2776
2777int ocfs2_nfs_sync_lock(struct ocfs2_super *osb, int ex)
2778{
2779 int status;
2780 struct ocfs2_lock_res *lockres = &osb->osb_nfs_sync_lockres;
2781
2782 if (ocfs2_is_hard_readonly(osb))
2783 return -EROFS;
2784
2785 if (ocfs2_mount_local(osb))
2786 return 0;
2787
2788 status = ocfs2_cluster_lock(osb, lockres, ex ? LKM_EXMODE : LKM_PRMODE,
2789 0, 0);
2790 if (status < 0)
2791 mlog(ML_ERROR, "lock on nfs sync lock failed %d\n", status);
2792
2793 return status;
2794}
2795
2796void ocfs2_nfs_sync_unlock(struct ocfs2_super *osb, int ex)
2797{
2798 struct ocfs2_lock_res *lockres = &osb->osb_nfs_sync_lockres;
2799
2800 if (!ocfs2_mount_local(osb))
2801 ocfs2_cluster_unlock(osb, lockres,
2802 ex ? LKM_EXMODE : LKM_PRMODE);
2803}
2804
2805int ocfs2_trim_fs_lock(struct ocfs2_super *osb,
2806 struct ocfs2_trim_fs_info *info, int trylock)
2807{
2808 int status;
2809 struct ocfs2_trim_fs_lvb *lvb;
2810 struct ocfs2_lock_res *lockres = &osb->osb_trim_fs_lockres;
2811
2812 if (info)
2813 info->tf_valid = 0;
2814
2815 if (ocfs2_is_hard_readonly(osb))
2816 return -EROFS;
2817
2818 if (ocfs2_mount_local(osb))
2819 return 0;
2820
2821 status = ocfs2_cluster_lock(osb, lockres, DLM_LOCK_EX,
2822 trylock ? DLM_LKF_NOQUEUE : 0, 0);
2823 if (status < 0) {
2824 if (status != -EAGAIN)
2825 mlog_errno(status);
2826 return status;
2827 }
2828
2829 if (info) {
2830 lvb = ocfs2_dlm_lvb(&lockres->l_lksb);
2831 if (ocfs2_dlm_lvb_valid(&lockres->l_lksb) &&
2832 lvb->lvb_version == OCFS2_TRIMFS_LVB_VERSION) {
2833 info->tf_valid = 1;
2834 info->tf_success = lvb->lvb_success;
2835 info->tf_nodenum = be32_to_cpu(lvb->lvb_nodenum);
2836 info->tf_start = be64_to_cpu(lvb->lvb_start);
2837 info->tf_len = be64_to_cpu(lvb->lvb_len);
2838 info->tf_minlen = be64_to_cpu(lvb->lvb_minlen);
2839 info->tf_trimlen = be64_to_cpu(lvb->lvb_trimlen);
2840 }
2841 }
2842
2843 return status;
2844}
2845
2846void ocfs2_trim_fs_unlock(struct ocfs2_super *osb,
2847 struct ocfs2_trim_fs_info *info)
2848{
2849 struct ocfs2_trim_fs_lvb *lvb;
2850 struct ocfs2_lock_res *lockres = &osb->osb_trim_fs_lockres;
2851
2852 if (ocfs2_mount_local(osb))
2853 return;
2854
2855 if (info) {
2856 lvb = ocfs2_dlm_lvb(&lockres->l_lksb);
2857 lvb->lvb_version = OCFS2_TRIMFS_LVB_VERSION;
2858 lvb->lvb_success = info->tf_success;
2859 lvb->lvb_nodenum = cpu_to_be32(info->tf_nodenum);
2860 lvb->lvb_start = cpu_to_be64(info->tf_start);
2861 lvb->lvb_len = cpu_to_be64(info->tf_len);
2862 lvb->lvb_minlen = cpu_to_be64(info->tf_minlen);
2863 lvb->lvb_trimlen = cpu_to_be64(info->tf_trimlen);
2864 }
2865
2866 ocfs2_cluster_unlock(osb, lockres, DLM_LOCK_EX);
2867}
2868
2869int ocfs2_dentry_lock(struct dentry *dentry, int ex)
2870{
2871 int ret;
2872 int level = ex ? DLM_LOCK_EX : DLM_LOCK_PR;
2873 struct ocfs2_dentry_lock *dl = dentry->d_fsdata;
2874 struct ocfs2_super *osb = OCFS2_SB(dentry->d_sb);
2875
2876 BUG_ON(!dl);
2877
2878 if (ocfs2_is_hard_readonly(osb)) {
2879 if (ex)
2880 return -EROFS;
2881 return 0;
2882 }
2883
2884 if (ocfs2_mount_local(osb))
2885 return 0;
2886
2887 ret = ocfs2_cluster_lock(osb, &dl->dl_lockres, level, 0, 0);
2888 if (ret < 0)
2889 mlog_errno(ret);
2890
2891 return ret;
2892}
2893
2894void ocfs2_dentry_unlock(struct dentry *dentry, int ex)
2895{
2896 int level = ex ? DLM_LOCK_EX : DLM_LOCK_PR;
2897 struct ocfs2_dentry_lock *dl = dentry->d_fsdata;
2898 struct ocfs2_super *osb = OCFS2_SB(dentry->d_sb);
2899
2900 if (!ocfs2_is_hard_readonly(osb) && !ocfs2_mount_local(osb))
2901 ocfs2_cluster_unlock(osb, &dl->dl_lockres, level);
2902}
2903
2904/* Reference counting of the dlm debug structure. We want this because
2905 * open references on the debug inodes can live on after a mount, so
2906 * we can't rely on the ocfs2_super to always exist. */
2907static void ocfs2_dlm_debug_free(struct kref *kref)
2908{
2909 struct ocfs2_dlm_debug *dlm_debug;
2910
2911 dlm_debug = container_of(kref, struct ocfs2_dlm_debug, d_refcnt);
2912
2913 kfree(dlm_debug);
2914}
2915
2916void ocfs2_put_dlm_debug(struct ocfs2_dlm_debug *dlm_debug)
2917{
2918 if (dlm_debug)
2919 kref_put(&dlm_debug->d_refcnt, ocfs2_dlm_debug_free);
2920}
2921
2922static void ocfs2_get_dlm_debug(struct ocfs2_dlm_debug *debug)
2923{
2924 kref_get(&debug->d_refcnt);
2925}
2926
2927struct ocfs2_dlm_debug *ocfs2_new_dlm_debug(void)
2928{
2929 struct ocfs2_dlm_debug *dlm_debug;
2930
2931 dlm_debug = kmalloc(sizeof(struct ocfs2_dlm_debug), GFP_KERNEL);
2932 if (!dlm_debug) {
2933 mlog_errno(-ENOMEM);
2934 goto out;
2935 }
2936
2937 kref_init(&dlm_debug->d_refcnt);
2938 INIT_LIST_HEAD(&dlm_debug->d_lockres_tracking);
2939 dlm_debug->d_locking_state = NULL;
2940out:
2941 return dlm_debug;
2942}
2943
2944/* Access to this is arbitrated for us via seq_file->sem. */
2945struct ocfs2_dlm_seq_priv {
2946 struct ocfs2_dlm_debug *p_dlm_debug;
2947 struct ocfs2_lock_res p_iter_res;
2948 struct ocfs2_lock_res p_tmp_res;
2949};
2950
2951static struct ocfs2_lock_res *ocfs2_dlm_next_res(struct ocfs2_lock_res *start,
2952 struct ocfs2_dlm_seq_priv *priv)
2953{
2954 struct ocfs2_lock_res *iter, *ret = NULL;
2955 struct ocfs2_dlm_debug *dlm_debug = priv->p_dlm_debug;
2956
2957 assert_spin_locked(&ocfs2_dlm_tracking_lock);
2958
2959 list_for_each_entry(iter, &start->l_debug_list, l_debug_list) {
2960 /* discover the head of the list */
2961 if (&iter->l_debug_list == &dlm_debug->d_lockres_tracking) {
2962 mlog(0, "End of list found, %p\n", ret);
2963 break;
2964 }
2965
2966 /* We track our "dummy" iteration lockres' by a NULL
2967 * l_ops field. */
2968 if (iter->l_ops != NULL) {
2969 ret = iter;
2970 break;
2971 }
2972 }
2973
2974 return ret;
2975}
2976
2977static void *ocfs2_dlm_seq_start(struct seq_file *m, loff_t *pos)
2978{
2979 struct ocfs2_dlm_seq_priv *priv = m->private;
2980 struct ocfs2_lock_res *iter;
2981
2982 spin_lock(&ocfs2_dlm_tracking_lock);
2983 iter = ocfs2_dlm_next_res(&priv->p_iter_res, priv);
2984 if (iter) {
2985 /* Since lockres' have the lifetime of their container
2986 * (which can be inodes, ocfs2_supers, etc) we want to
2987 * copy this out to a temporary lockres while still
2988 * under the spinlock. Obviously after this we can't
2989 * trust any pointers on the copy returned, but that's
2990 * ok as the information we want isn't typically held
2991 * in them. */
2992 priv->p_tmp_res = *iter;
2993 iter = &priv->p_tmp_res;
2994 }
2995 spin_unlock(&ocfs2_dlm_tracking_lock);
2996
2997 return iter;
2998}
2999
3000static void ocfs2_dlm_seq_stop(struct seq_file *m, void *v)
3001{
3002}
3003
3004static void *ocfs2_dlm_seq_next(struct seq_file *m, void *v, loff_t *pos)
3005{
3006 struct ocfs2_dlm_seq_priv *priv = m->private;
3007 struct ocfs2_lock_res *iter = v;
3008 struct ocfs2_lock_res *dummy = &priv->p_iter_res;
3009
3010 spin_lock(&ocfs2_dlm_tracking_lock);
3011 iter = ocfs2_dlm_next_res(iter, priv);
3012 list_del_init(&dummy->l_debug_list);
3013 if (iter) {
3014 list_add(&dummy->l_debug_list, &iter->l_debug_list);
3015 priv->p_tmp_res = *iter;
3016 iter = &priv->p_tmp_res;
3017 }
3018 spin_unlock(&ocfs2_dlm_tracking_lock);
3019
3020 return iter;
3021}
3022
3023/*
3024 * Version is used by debugfs.ocfs2 to determine the format being used
3025 *
3026 * New in version 2
3027 * - Lock stats printed
3028 * New in version 3
3029 * - Max time in lock stats is in usecs (instead of nsecs)
3030 */
3031#define OCFS2_DLM_DEBUG_STR_VERSION 3
3032static int ocfs2_dlm_seq_show(struct seq_file *m, void *v)
3033{
3034 int i;
3035 char *lvb;
3036 struct ocfs2_lock_res *lockres = v;
3037
3038 if (!lockres)
3039 return -EINVAL;
3040
3041 seq_printf(m, "0x%x\t", OCFS2_DLM_DEBUG_STR_VERSION);
3042
3043 if (lockres->l_type == OCFS2_LOCK_TYPE_DENTRY)
3044 seq_printf(m, "%.*s%08x\t", OCFS2_DENTRY_LOCK_INO_START - 1,
3045 lockres->l_name,
3046 (unsigned int)ocfs2_get_dentry_lock_ino(lockres));
3047 else
3048 seq_printf(m, "%.*s\t", OCFS2_LOCK_ID_MAX_LEN, lockres->l_name);
3049
3050 seq_printf(m, "%d\t"
3051 "0x%lx\t"
3052 "0x%x\t"
3053 "0x%x\t"
3054 "%u\t"
3055 "%u\t"
3056 "%d\t"
3057 "%d\t",
3058 lockres->l_level,
3059 lockres->l_flags,
3060 lockres->l_action,
3061 lockres->l_unlock_action,
3062 lockres->l_ro_holders,
3063 lockres->l_ex_holders,
3064 lockres->l_requested,
3065 lockres->l_blocking);
3066
3067 /* Dump the raw LVB */
3068 lvb = ocfs2_dlm_lvb(&lockres->l_lksb);
3069 for(i = 0; i < DLM_LVB_LEN; i++)
3070 seq_printf(m, "0x%x\t", lvb[i]);
3071
3072#ifdef CONFIG_OCFS2_FS_STATS
3073# define lock_num_prmode(_l) ((_l)->l_lock_prmode.ls_gets)
3074# define lock_num_exmode(_l) ((_l)->l_lock_exmode.ls_gets)
3075# define lock_num_prmode_failed(_l) ((_l)->l_lock_prmode.ls_fail)
3076# define lock_num_exmode_failed(_l) ((_l)->l_lock_exmode.ls_fail)
3077# define lock_total_prmode(_l) ((_l)->l_lock_prmode.ls_total)
3078# define lock_total_exmode(_l) ((_l)->l_lock_exmode.ls_total)
3079# define lock_max_prmode(_l) ((_l)->l_lock_prmode.ls_max)
3080# define lock_max_exmode(_l) ((_l)->l_lock_exmode.ls_max)
3081# define lock_refresh(_l) ((_l)->l_lock_refresh)
3082#else
3083# define lock_num_prmode(_l) (0)
3084# define lock_num_exmode(_l) (0)
3085# define lock_num_prmode_failed(_l) (0)
3086# define lock_num_exmode_failed(_l) (0)
3087# define lock_total_prmode(_l) (0ULL)
3088# define lock_total_exmode(_l) (0ULL)
3089# define lock_max_prmode(_l) (0)
3090# define lock_max_exmode(_l) (0)
3091# define lock_refresh(_l) (0)
3092#endif
3093 /* The following seq_print was added in version 2 of this output */
3094 seq_printf(m, "%u\t"
3095 "%u\t"
3096 "%u\t"
3097 "%u\t"
3098 "%llu\t"
3099 "%llu\t"
3100 "%u\t"
3101 "%u\t"
3102 "%u\t",
3103 lock_num_prmode(lockres),
3104 lock_num_exmode(lockres),
3105 lock_num_prmode_failed(lockres),
3106 lock_num_exmode_failed(lockres),
3107 lock_total_prmode(lockres),
3108 lock_total_exmode(lockres),
3109 lock_max_prmode(lockres),
3110 lock_max_exmode(lockres),
3111 lock_refresh(lockres));
3112
3113 /* End the line */
3114 seq_printf(m, "\n");
3115 return 0;
3116}
3117
3118static const struct seq_operations ocfs2_dlm_seq_ops = {
3119 .start = ocfs2_dlm_seq_start,
3120 .stop = ocfs2_dlm_seq_stop,
3121 .next = ocfs2_dlm_seq_next,
3122 .show = ocfs2_dlm_seq_show,
3123};
3124
3125static int ocfs2_dlm_debug_release(struct inode *inode, struct file *file)
3126{
3127 struct seq_file *seq = file->private_data;
3128 struct ocfs2_dlm_seq_priv *priv = seq->private;
3129 struct ocfs2_lock_res *res = &priv->p_iter_res;
3130
3131 ocfs2_remove_lockres_tracking(res);
3132 ocfs2_put_dlm_debug(priv->p_dlm_debug);
3133 return seq_release_private(inode, file);
3134}
3135
3136static int ocfs2_dlm_debug_open(struct inode *inode, struct file *file)
3137{
3138 struct ocfs2_dlm_seq_priv *priv;
3139 struct ocfs2_super *osb;
3140
3141 priv = __seq_open_private(file, &ocfs2_dlm_seq_ops, sizeof(*priv));
3142 if (!priv) {
3143 mlog_errno(-ENOMEM);
3144 return -ENOMEM;
3145 }
3146
3147 osb = inode->i_private;
3148 ocfs2_get_dlm_debug(osb->osb_dlm_debug);
3149 priv->p_dlm_debug = osb->osb_dlm_debug;
3150 INIT_LIST_HEAD(&priv->p_iter_res.l_debug_list);
3151
3152 ocfs2_add_lockres_tracking(&priv->p_iter_res,
3153 priv->p_dlm_debug);
3154
3155 return 0;
3156}
3157
3158static const struct file_operations ocfs2_dlm_debug_fops = {
3159 .open = ocfs2_dlm_debug_open,
3160 .release = ocfs2_dlm_debug_release,
3161 .read = seq_read,
3162 .llseek = seq_lseek,
3163};
3164
3165static int ocfs2_dlm_init_debug(struct ocfs2_super *osb)
3166{
3167 int ret = 0;
3168 struct ocfs2_dlm_debug *dlm_debug = osb->osb_dlm_debug;
3169
3170 dlm_debug->d_locking_state = debugfs_create_file("locking_state",
3171 S_IFREG|S_IRUSR,
3172 osb->osb_debug_root,
3173 osb,
3174 &ocfs2_dlm_debug_fops);
3175 if (!dlm_debug->d_locking_state) {
3176 ret = -EINVAL;
3177 mlog(ML_ERROR,
3178 "Unable to create locking state debugfs file.\n");
3179 goto out;
3180 }
3181
3182 ocfs2_get_dlm_debug(dlm_debug);
3183out:
3184 return ret;
3185}
3186
3187static void ocfs2_dlm_shutdown_debug(struct ocfs2_super *osb)
3188{
3189 struct ocfs2_dlm_debug *dlm_debug = osb->osb_dlm_debug;
3190
3191 if (dlm_debug) {
3192 debugfs_remove(dlm_debug->d_locking_state);
3193 ocfs2_put_dlm_debug(dlm_debug);
3194 }
3195}
3196
3197int ocfs2_dlm_init(struct ocfs2_super *osb)
3198{
3199 int status = 0;
3200 struct ocfs2_cluster_connection *conn = NULL;
3201
3202 if (ocfs2_mount_local(osb)) {
3203 osb->node_num = 0;
3204 goto local;
3205 }
3206
3207 status = ocfs2_dlm_init_debug(osb);
3208 if (status < 0) {
3209 mlog_errno(status);
3210 goto bail;
3211 }
3212
3213 /* launch downconvert thread */
3214 osb->dc_task = kthread_run(ocfs2_downconvert_thread, osb, "ocfs2dc-%s",
3215 osb->uuid_str);
3216 if (IS_ERR(osb->dc_task)) {
3217 status = PTR_ERR(osb->dc_task);
3218 osb->dc_task = NULL;
3219 mlog_errno(status);
3220 goto bail;
3221 }
3222
3223 /* for now, uuid == domain */
3224 status = ocfs2_cluster_connect(osb->osb_cluster_stack,
3225 osb->osb_cluster_name,
3226 strlen(osb->osb_cluster_name),
3227 osb->uuid_str,
3228 strlen(osb->uuid_str),
3229 &lproto, ocfs2_do_node_down, osb,
3230 &conn);
3231 if (status) {
3232 mlog_errno(status);
3233 goto bail;
3234 }
3235
3236 status = ocfs2_cluster_this_node(conn, &osb->node_num);
3237 if (status < 0) {
3238 mlog_errno(status);
3239 mlog(ML_ERROR,
3240 "could not find this host's node number\n");
3241 ocfs2_cluster_disconnect(conn, 0);
3242 goto bail;
3243 }
3244
3245local:
3246 ocfs2_super_lock_res_init(&osb->osb_super_lockres, osb);
3247 ocfs2_rename_lock_res_init(&osb->osb_rename_lockres, osb);
3248 ocfs2_nfs_sync_lock_res_init(&osb->osb_nfs_sync_lockres, osb);
3249 ocfs2_orphan_scan_lock_res_init(&osb->osb_orphan_scan.os_lockres, osb);
3250
3251 osb->cconn = conn;
3252bail:
3253 if (status < 0) {
3254 ocfs2_dlm_shutdown_debug(osb);
3255 if (osb->dc_task)
3256 kthread_stop(osb->dc_task);
3257 }
3258
3259 return status;
3260}
3261
3262void ocfs2_dlm_shutdown(struct ocfs2_super *osb,
3263 int hangup_pending)
3264{
3265 ocfs2_drop_osb_locks(osb);
3266
3267 /*
3268 * Now that we have dropped all locks and ocfs2_dismount_volume()
3269 * has disabled recovery, the DLM won't be talking to us. It's
3270 * safe to tear things down before disconnecting the cluster.
3271 */
3272
3273 if (osb->dc_task) {
3274 kthread_stop(osb->dc_task);
3275 osb->dc_task = NULL;
3276 }
3277
3278 ocfs2_lock_res_free(&osb->osb_super_lockres);
3279 ocfs2_lock_res_free(&osb->osb_rename_lockres);
3280 ocfs2_lock_res_free(&osb->osb_nfs_sync_lockres);
3281 ocfs2_lock_res_free(&osb->osb_orphan_scan.os_lockres);
3282
3283 ocfs2_cluster_disconnect(osb->cconn, hangup_pending);
3284 osb->cconn = NULL;
3285
3286 ocfs2_dlm_shutdown_debug(osb);
3287}
3288
3289static int ocfs2_drop_lock(struct ocfs2_super *osb,
3290 struct ocfs2_lock_res *lockres)
3291{
3292 int ret;
3293 unsigned long flags;
3294 u32 lkm_flags = 0;
3295
3296 /* We didn't get anywhere near actually using this lockres. */
3297 if (!(lockres->l_flags & OCFS2_LOCK_INITIALIZED))
3298 goto out;
3299
3300 if (lockres->l_ops->flags & LOCK_TYPE_USES_LVB)
3301 lkm_flags |= DLM_LKF_VALBLK;
3302
3303 spin_lock_irqsave(&lockres->l_lock, flags);
3304
3305 mlog_bug_on_msg(!(lockres->l_flags & OCFS2_LOCK_FREEING),
3306 "lockres %s, flags 0x%lx\n",
3307 lockres->l_name, lockres->l_flags);
3308
3309 while (lockres->l_flags & OCFS2_LOCK_BUSY) {
3310 mlog(0, "waiting on busy lock \"%s\": flags = %lx, action = "
3311 "%u, unlock_action = %u\n",
3312 lockres->l_name, lockres->l_flags, lockres->l_action,
3313 lockres->l_unlock_action);
3314
3315 spin_unlock_irqrestore(&lockres->l_lock, flags);
3316
3317 /* XXX: Today we just wait on any busy
3318 * locks... Perhaps we need to cancel converts in the
3319 * future? */
3320 ocfs2_wait_on_busy_lock(lockres);
3321
3322 spin_lock_irqsave(&lockres->l_lock, flags);
3323 }
3324
3325 if (lockres->l_ops->flags & LOCK_TYPE_USES_LVB) {
3326 if (lockres->l_flags & OCFS2_LOCK_ATTACHED &&
3327 lockres->l_level == DLM_LOCK_EX &&
3328 !(lockres->l_flags & OCFS2_LOCK_NEEDS_REFRESH))
3329 lockres->l_ops->set_lvb(lockres);
3330 }
3331
3332 if (lockres->l_flags & OCFS2_LOCK_BUSY)
3333 mlog(ML_ERROR, "destroying busy lock: \"%s\"\n",
3334 lockres->l_name);
3335 if (lockres->l_flags & OCFS2_LOCK_BLOCKED)
3336 mlog(0, "destroying blocked lock: \"%s\"\n", lockres->l_name);
3337
3338 if (!(lockres->l_flags & OCFS2_LOCK_ATTACHED)) {
3339 spin_unlock_irqrestore(&lockres->l_lock, flags);
3340 goto out;
3341 }
3342
3343 lockres_clear_flags(lockres, OCFS2_LOCK_ATTACHED);
3344
3345 /* make sure we never get here while waiting for an ast to
3346 * fire. */
3347 BUG_ON(lockres->l_action != OCFS2_AST_INVALID);
3348
3349 /* is this necessary? */
3350 lockres_or_flags(lockres, OCFS2_LOCK_BUSY);
3351 lockres->l_unlock_action = OCFS2_UNLOCK_DROP_LOCK;
3352 spin_unlock_irqrestore(&lockres->l_lock, flags);
3353
3354 mlog(0, "lock %s\n", lockres->l_name);
3355
3356 ret = ocfs2_dlm_unlock(osb->cconn, &lockres->l_lksb, lkm_flags);
3357 if (ret) {
3358 ocfs2_log_dlm_error("ocfs2_dlm_unlock", ret, lockres);
3359 mlog(ML_ERROR, "lockres flags: %lu\n", lockres->l_flags);
3360 ocfs2_dlm_dump_lksb(&lockres->l_lksb);
3361 BUG();
3362 }
3363 mlog(0, "lock %s, successful return from ocfs2_dlm_unlock\n",
3364 lockres->l_name);
3365
3366 ocfs2_wait_on_busy_lock(lockres);
3367out:
3368 return 0;
3369}
3370
3371static void ocfs2_process_blocked_lock(struct ocfs2_super *osb,
3372 struct ocfs2_lock_res *lockres);
3373
3374/* Mark the lockres as being dropped. It will no longer be
3375 * queued if blocking, but we still may have to wait on it
3376 * being dequeued from the downconvert thread before we can consider
3377 * it safe to drop.
3378 *
3379 * You can *not* attempt to call cluster_lock on this lockres anymore. */
3380void ocfs2_mark_lockres_freeing(struct ocfs2_super *osb,
3381 struct ocfs2_lock_res *lockres)
3382{
3383 int status;
3384 struct ocfs2_mask_waiter mw;
3385 unsigned long flags, flags2;
3386
3387 ocfs2_init_mask_waiter(&mw);
3388
3389 spin_lock_irqsave(&lockres->l_lock, flags);
3390 lockres->l_flags |= OCFS2_LOCK_FREEING;
3391 if (lockres->l_flags & OCFS2_LOCK_QUEUED && current == osb->dc_task) {
3392 /*
3393 * We know the downconvert is queued but not in progress
3394 * because we are the downconvert thread and processing
3395 * different lock. So we can just remove the lock from the
3396 * queue. This is not only an optimization but also a way
3397 * to avoid the following deadlock:
3398 * ocfs2_dentry_post_unlock()
3399 * ocfs2_dentry_lock_put()
3400 * ocfs2_drop_dentry_lock()
3401 * iput()
3402 * ocfs2_evict_inode()
3403 * ocfs2_clear_inode()
3404 * ocfs2_mark_lockres_freeing()
3405 * ... blocks waiting for OCFS2_LOCK_QUEUED
3406 * since we are the downconvert thread which
3407 * should clear the flag.
3408 */
3409 spin_unlock_irqrestore(&lockres->l_lock, flags);
3410 spin_lock_irqsave(&osb->dc_task_lock, flags2);
3411 list_del_init(&lockres->l_blocked_list);
3412 osb->blocked_lock_count--;
3413 spin_unlock_irqrestore(&osb->dc_task_lock, flags2);
3414 /*
3415 * Warn if we recurse into another post_unlock call. Strictly
3416 * speaking it isn't a problem but we need to be careful if
3417 * that happens (stack overflow, deadlocks, ...) so warn if
3418 * ocfs2 grows a path for which this can happen.
3419 */
3420 WARN_ON_ONCE(lockres->l_ops->post_unlock);
3421 /* Since the lock is freeing we don't do much in the fn below */
3422 ocfs2_process_blocked_lock(osb, lockres);
3423 return;
3424 }
3425 while (lockres->l_flags & OCFS2_LOCK_QUEUED) {
3426 lockres_add_mask_waiter(lockres, &mw, OCFS2_LOCK_QUEUED, 0);
3427 spin_unlock_irqrestore(&lockres->l_lock, flags);
3428
3429 mlog(0, "Waiting on lockres %s\n", lockres->l_name);
3430
3431 status = ocfs2_wait_for_mask(&mw);
3432 if (status)
3433 mlog_errno(status);
3434
3435 spin_lock_irqsave(&lockres->l_lock, flags);
3436 }
3437 spin_unlock_irqrestore(&lockres->l_lock, flags);
3438}
3439
3440void ocfs2_simple_drop_lockres(struct ocfs2_super *osb,
3441 struct ocfs2_lock_res *lockres)
3442{
3443 int ret;
3444
3445 ocfs2_mark_lockres_freeing(osb, lockres);
3446 ret = ocfs2_drop_lock(osb, lockres);
3447 if (ret)
3448 mlog_errno(ret);
3449}
3450
3451static void ocfs2_drop_osb_locks(struct ocfs2_super *osb)
3452{
3453 ocfs2_simple_drop_lockres(osb, &osb->osb_super_lockres);
3454 ocfs2_simple_drop_lockres(osb, &osb->osb_rename_lockres);
3455 ocfs2_simple_drop_lockres(osb, &osb->osb_nfs_sync_lockres);
3456 ocfs2_simple_drop_lockres(osb, &osb->osb_orphan_scan.os_lockres);
3457}
3458
3459int ocfs2_drop_inode_locks(struct inode *inode)
3460{
3461 int status, err;
3462
3463 /* No need to call ocfs2_mark_lockres_freeing here -
3464 * ocfs2_clear_inode has done it for us. */
3465
3466 err = ocfs2_drop_lock(OCFS2_SB(inode->i_sb),
3467 &OCFS2_I(inode)->ip_open_lockres);
3468 if (err < 0)
3469 mlog_errno(err);
3470
3471 status = err;
3472
3473 err = ocfs2_drop_lock(OCFS2_SB(inode->i_sb),
3474 &OCFS2_I(inode)->ip_inode_lockres);
3475 if (err < 0)
3476 mlog_errno(err);
3477 if (err < 0 && !status)
3478 status = err;
3479
3480 err = ocfs2_drop_lock(OCFS2_SB(inode->i_sb),
3481 &OCFS2_I(inode)->ip_rw_lockres);
3482 if (err < 0)
3483 mlog_errno(err);
3484 if (err < 0 && !status)
3485 status = err;
3486
3487 return status;
3488}
3489
3490static unsigned int ocfs2_prepare_downconvert(struct ocfs2_lock_res *lockres,
3491 int new_level)
3492{
3493 assert_spin_locked(&lockres->l_lock);
3494
3495 BUG_ON(lockres->l_blocking <= DLM_LOCK_NL);
3496
3497 if (lockres->l_level <= new_level) {
3498 mlog(ML_ERROR, "lockres %s, lvl %d <= %d, blcklst %d, mask %d, "
3499 "type %d, flags 0x%lx, hold %d %d, act %d %d, req %d, "
3500 "block %d, pgen %d\n", lockres->l_name, lockres->l_level,
3501 new_level, list_empty(&lockres->l_blocked_list),
3502 list_empty(&lockres->l_mask_waiters), lockres->l_type,
3503 lockres->l_flags, lockres->l_ro_holders,
3504 lockres->l_ex_holders, lockres->l_action,
3505 lockres->l_unlock_action, lockres->l_requested,
3506 lockres->l_blocking, lockres->l_pending_gen);
3507 BUG();
3508 }
3509
3510 mlog(ML_BASTS, "lockres %s, level %d => %d, blocking %d\n",
3511 lockres->l_name, lockres->l_level, new_level, lockres->l_blocking);
3512
3513 lockres->l_action = OCFS2_AST_DOWNCONVERT;
3514 lockres->l_requested = new_level;
3515 lockres_or_flags(lockres, OCFS2_LOCK_BUSY);
3516 return lockres_set_pending(lockres);
3517}
3518
3519static int ocfs2_downconvert_lock(struct ocfs2_super *osb,
3520 struct ocfs2_lock_res *lockres,
3521 int new_level,
3522 int lvb,
3523 unsigned int generation)
3524{
3525 int ret;
3526 u32 dlm_flags = DLM_LKF_CONVERT;
3527
3528 mlog(ML_BASTS, "lockres %s, level %d => %d\n", lockres->l_name,
3529 lockres->l_level, new_level);
3530
3531 /*
3532 * On DLM_LKF_VALBLK, fsdlm behaves differently with o2cb. It always
3533 * expects DLM_LKF_VALBLK being set if the LKB has LVB, so that
3534 * we can recover correctly from node failure. Otherwise, we may get
3535 * invalid LVB in LKB, but without DLM_SBF_VALNOTVALID being set.
3536 */
3537 if (!ocfs2_is_o2cb_active() &&
3538 lockres->l_ops->flags & LOCK_TYPE_USES_LVB)
3539 lvb = 1;
3540
3541 if (lvb)
3542 dlm_flags |= DLM_LKF_VALBLK;
3543
3544 ret = ocfs2_dlm_lock(osb->cconn,
3545 new_level,
3546 &lockres->l_lksb,
3547 dlm_flags,
3548 lockres->l_name,
3549 OCFS2_LOCK_ID_MAX_LEN - 1);
3550 lockres_clear_pending(lockres, generation, osb);
3551 if (ret) {
3552 ocfs2_log_dlm_error("ocfs2_dlm_lock", ret, lockres);
3553 ocfs2_recover_from_dlm_error(lockres, 1);
3554 goto bail;
3555 }
3556
3557 ret = 0;
3558bail:
3559 return ret;
3560}
3561
3562/* returns 1 when the caller should unlock and call ocfs2_dlm_unlock */
3563static int ocfs2_prepare_cancel_convert(struct ocfs2_super *osb,
3564 struct ocfs2_lock_res *lockres)
3565{
3566 assert_spin_locked(&lockres->l_lock);
3567
3568 if (lockres->l_unlock_action == OCFS2_UNLOCK_CANCEL_CONVERT) {
3569 /* If we're already trying to cancel a lock conversion
3570 * then just drop the spinlock and allow the caller to
3571 * requeue this lock. */
3572 mlog(ML_BASTS, "lockres %s, skip convert\n", lockres->l_name);
3573 return 0;
3574 }
3575
3576 /* were we in a convert when we got the bast fire? */
3577 BUG_ON(lockres->l_action != OCFS2_AST_CONVERT &&
3578 lockres->l_action != OCFS2_AST_DOWNCONVERT);
3579 /* set things up for the unlockast to know to just
3580 * clear out the ast_action and unset busy, etc. */
3581 lockres->l_unlock_action = OCFS2_UNLOCK_CANCEL_CONVERT;
3582
3583 mlog_bug_on_msg(!(lockres->l_flags & OCFS2_LOCK_BUSY),
3584 "lock %s, invalid flags: 0x%lx\n",
3585 lockres->l_name, lockres->l_flags);
3586
3587 mlog(ML_BASTS, "lockres %s\n", lockres->l_name);
3588
3589 return 1;
3590}
3591
3592static int ocfs2_cancel_convert(struct ocfs2_super *osb,
3593 struct ocfs2_lock_res *lockres)
3594{
3595 int ret;
3596
3597 ret = ocfs2_dlm_unlock(osb->cconn, &lockres->l_lksb,
3598 DLM_LKF_CANCEL);
3599 if (ret) {
3600 ocfs2_log_dlm_error("ocfs2_dlm_unlock", ret, lockres);
3601 ocfs2_recover_from_dlm_error(lockres, 0);
3602 }
3603
3604 mlog(ML_BASTS, "lockres %s\n", lockres->l_name);
3605
3606 return ret;
3607}
3608
3609static int ocfs2_unblock_lock(struct ocfs2_super *osb,
3610 struct ocfs2_lock_res *lockres,
3611 struct ocfs2_unblock_ctl *ctl)
3612{
3613 unsigned long flags;
3614 int blocking;
3615 int new_level;
3616 int level;
3617 int ret = 0;
3618 int set_lvb = 0;
3619 unsigned int gen;
3620
3621 spin_lock_irqsave(&lockres->l_lock, flags);
3622
3623recheck:
3624 /*
3625 * Is it still blocking? If not, we have no more work to do.
3626 */
3627 if (!(lockres->l_flags & OCFS2_LOCK_BLOCKED)) {
3628 BUG_ON(lockres->l_blocking != DLM_LOCK_NL);
3629 spin_unlock_irqrestore(&lockres->l_lock, flags);
3630 ret = 0;
3631 goto leave;
3632 }
3633
3634 if (lockres->l_flags & OCFS2_LOCK_BUSY) {
3635 /* XXX
3636 * This is a *big* race. The OCFS2_LOCK_PENDING flag
3637 * exists entirely for one reason - another thread has set
3638 * OCFS2_LOCK_BUSY, but has *NOT* yet called dlm_lock().
3639 *
3640 * If we do ocfs2_cancel_convert() before the other thread
3641 * calls dlm_lock(), our cancel will do nothing. We will
3642 * get no ast, and we will have no way of knowing the
3643 * cancel failed. Meanwhile, the other thread will call
3644 * into dlm_lock() and wait...forever.
3645 *
3646 * Why forever? Because another node has asked for the
3647 * lock first; that's why we're here in unblock_lock().
3648 *
3649 * The solution is OCFS2_LOCK_PENDING. When PENDING is
3650 * set, we just requeue the unblock. Only when the other
3651 * thread has called dlm_lock() and cleared PENDING will
3652 * we then cancel their request.
3653 *
3654 * All callers of dlm_lock() must set OCFS2_DLM_PENDING
3655 * at the same time they set OCFS2_DLM_BUSY. They must
3656 * clear OCFS2_DLM_PENDING after dlm_lock() returns.
3657 */
3658 if (lockres->l_flags & OCFS2_LOCK_PENDING) {
3659 mlog(ML_BASTS, "lockres %s, ReQ: Pending\n",
3660 lockres->l_name);
3661 goto leave_requeue;
3662 }
3663
3664 ctl->requeue = 1;
3665 ret = ocfs2_prepare_cancel_convert(osb, lockres);
3666 spin_unlock_irqrestore(&lockres->l_lock, flags);
3667 if (ret) {
3668 ret = ocfs2_cancel_convert(osb, lockres);
3669 if (ret < 0)
3670 mlog_errno(ret);
3671 }
3672 goto leave;
3673 }
3674
3675 /*
3676 * This prevents livelocks. OCFS2_LOCK_UPCONVERT_FINISHING flag is
3677 * set when the ast is received for an upconvert just before the
3678 * OCFS2_LOCK_BUSY flag is cleared. Now if the fs received a bast
3679 * on the heels of the ast, we want to delay the downconvert just
3680 * enough to allow the up requestor to do its task. Because this
3681 * lock is in the blocked queue, the lock will be downconverted
3682 * as soon as the requestor is done with the lock.
3683 */
3684 if (lockres->l_flags & OCFS2_LOCK_UPCONVERT_FINISHING)
3685 goto leave_requeue;
3686
3687 /*
3688 * How can we block and yet be at NL? We were trying to upconvert
3689 * from NL and got canceled. The code comes back here, and now
3690 * we notice and clear BLOCKING.
3691 */
3692 if (lockres->l_level == DLM_LOCK_NL) {
3693 BUG_ON(lockres->l_ex_holders || lockres->l_ro_holders);
3694 mlog(ML_BASTS, "lockres %s, Aborting dc\n", lockres->l_name);
3695 lockres->l_blocking = DLM_LOCK_NL;
3696 lockres_clear_flags(lockres, OCFS2_LOCK_BLOCKED);
3697 spin_unlock_irqrestore(&lockres->l_lock, flags);
3698 goto leave;
3699 }
3700
3701 /* if we're blocking an exclusive and we have *any* holders,
3702 * then requeue. */
3703 if ((lockres->l_blocking == DLM_LOCK_EX)
3704 && (lockres->l_ex_holders || lockres->l_ro_holders)) {
3705 mlog(ML_BASTS, "lockres %s, ReQ: EX/PR Holders %u,%u\n",
3706 lockres->l_name, lockres->l_ex_holders,
3707 lockres->l_ro_holders);
3708 goto leave_requeue;
3709 }
3710
3711 /* If it's a PR we're blocking, then only
3712 * requeue if we've got any EX holders */
3713 if (lockres->l_blocking == DLM_LOCK_PR &&
3714 lockres->l_ex_holders) {
3715 mlog(ML_BASTS, "lockres %s, ReQ: EX Holders %u\n",
3716 lockres->l_name, lockres->l_ex_holders);
3717 goto leave_requeue;
3718 }
3719
3720 /*
3721 * Can we get a lock in this state if the holder counts are
3722 * zero? The meta data unblock code used to check this.
3723 */
3724 if ((lockres->l_ops->flags & LOCK_TYPE_REQUIRES_REFRESH)
3725 && (lockres->l_flags & OCFS2_LOCK_REFRESHING)) {
3726 mlog(ML_BASTS, "lockres %s, ReQ: Lock Refreshing\n",
3727 lockres->l_name);
3728 goto leave_requeue;
3729 }
3730
3731 new_level = ocfs2_highest_compat_lock_level(lockres->l_blocking);
3732
3733 if (lockres->l_ops->check_downconvert
3734 && !lockres->l_ops->check_downconvert(lockres, new_level)) {
3735 mlog(ML_BASTS, "lockres %s, ReQ: Checkpointing\n",
3736 lockres->l_name);
3737 goto leave_requeue;
3738 }
3739
3740 /* If we get here, then we know that there are no more
3741 * incompatible holders (and anyone asking for an incompatible
3742 * lock is blocked). We can now downconvert the lock */
3743 if (!lockres->l_ops->downconvert_worker)
3744 goto downconvert;
3745
3746 /* Some lockres types want to do a bit of work before
3747 * downconverting a lock. Allow that here. The worker function
3748 * may sleep, so we save off a copy of what we're blocking as
3749 * it may change while we're not holding the spin lock. */
3750 blocking = lockres->l_blocking;
3751 level = lockres->l_level;
3752 spin_unlock_irqrestore(&lockres->l_lock, flags);
3753
3754 ctl->unblock_action = lockres->l_ops->downconvert_worker(lockres, blocking);
3755
3756 if (ctl->unblock_action == UNBLOCK_STOP_POST) {
3757 mlog(ML_BASTS, "lockres %s, UNBLOCK_STOP_POST\n",
3758 lockres->l_name);
3759 goto leave;
3760 }
3761
3762 spin_lock_irqsave(&lockres->l_lock, flags);
3763 if ((blocking != lockres->l_blocking) || (level != lockres->l_level)) {
3764 /* If this changed underneath us, then we can't drop
3765 * it just yet. */
3766 mlog(ML_BASTS, "lockres %s, block=%d:%d, level=%d:%d, "
3767 "Recheck\n", lockres->l_name, blocking,
3768 lockres->l_blocking, level, lockres->l_level);
3769 goto recheck;
3770 }
3771
3772downconvert:
3773 ctl->requeue = 0;
3774
3775 if (lockres->l_ops->flags & LOCK_TYPE_USES_LVB) {
3776 if (lockres->l_level == DLM_LOCK_EX)
3777 set_lvb = 1;
3778
3779 /*
3780 * We only set the lvb if the lock has been fully
3781 * refreshed - otherwise we risk setting stale
3782 * data. Otherwise, there's no need to actually clear
3783 * out the lvb here as it's value is still valid.
3784 */
3785 if (set_lvb && !(lockres->l_flags & OCFS2_LOCK_NEEDS_REFRESH))
3786 lockres->l_ops->set_lvb(lockres);
3787 }
3788
3789 gen = ocfs2_prepare_downconvert(lockres, new_level);
3790 spin_unlock_irqrestore(&lockres->l_lock, flags);
3791 ret = ocfs2_downconvert_lock(osb, lockres, new_level, set_lvb,
3792 gen);
3793
3794leave:
3795 if (ret)
3796 mlog_errno(ret);
3797 return ret;
3798
3799leave_requeue:
3800 spin_unlock_irqrestore(&lockres->l_lock, flags);
3801 ctl->requeue = 1;
3802
3803 return 0;
3804}
3805
3806static int ocfs2_data_convert_worker(struct ocfs2_lock_res *lockres,
3807 int blocking)
3808{
3809 struct inode *inode;
3810 struct address_space *mapping;
3811 struct ocfs2_inode_info *oi;
3812
3813 inode = ocfs2_lock_res_inode(lockres);
3814 mapping = inode->i_mapping;
3815
3816 if (S_ISDIR(inode->i_mode)) {
3817 oi = OCFS2_I(inode);
3818 oi->ip_dir_lock_gen++;
3819 mlog(0, "generation: %u\n", oi->ip_dir_lock_gen);
3820 goto out;
3821 }
3822
3823 if (!S_ISREG(inode->i_mode))
3824 goto out;
3825
3826 /*
3827 * We need this before the filemap_fdatawrite() so that it can
3828 * transfer the dirty bit from the PTE to the
3829 * page. Unfortunately this means that even for EX->PR
3830 * downconverts, we'll lose our mappings and have to build
3831 * them up again.
3832 */
3833 unmap_mapping_range(mapping, 0, 0, 0);
3834
3835 if (filemap_fdatawrite(mapping)) {
3836 mlog(ML_ERROR, "Could not sync inode %llu for downconvert!",
3837 (unsigned long long)OCFS2_I(inode)->ip_blkno);
3838 }
3839 sync_mapping_buffers(mapping);
3840 if (blocking == DLM_LOCK_EX) {
3841 truncate_inode_pages(mapping, 0);
3842 } else {
3843 /* We only need to wait on the I/O if we're not also
3844 * truncating pages because truncate_inode_pages waits
3845 * for us above. We don't truncate pages if we're
3846 * blocking anything < EXMODE because we want to keep
3847 * them around in that case. */
3848 filemap_fdatawait(mapping);
3849 }
3850
3851 forget_all_cached_acls(inode);
3852
3853out:
3854 return UNBLOCK_CONTINUE;
3855}
3856
3857static int ocfs2_ci_checkpointed(struct ocfs2_caching_info *ci,
3858 struct ocfs2_lock_res *lockres,
3859 int new_level)
3860{
3861 int checkpointed = ocfs2_ci_fully_checkpointed(ci);
3862
3863 BUG_ON(new_level != DLM_LOCK_NL && new_level != DLM_LOCK_PR);
3864 BUG_ON(lockres->l_level != DLM_LOCK_EX && !checkpointed);
3865
3866 if (checkpointed)
3867 return 1;
3868
3869 ocfs2_start_checkpoint(OCFS2_SB(ocfs2_metadata_cache_get_super(ci)));
3870 return 0;
3871}
3872
3873static int ocfs2_check_meta_downconvert(struct ocfs2_lock_res *lockres,
3874 int new_level)
3875{
3876 struct inode *inode = ocfs2_lock_res_inode(lockres);
3877
3878 return ocfs2_ci_checkpointed(INODE_CACHE(inode), lockres, new_level);
3879}
3880
3881static void ocfs2_set_meta_lvb(struct ocfs2_lock_res *lockres)
3882{
3883 struct inode *inode = ocfs2_lock_res_inode(lockres);
3884
3885 __ocfs2_stuff_meta_lvb(inode);
3886}
3887
3888/*
3889 * Does the final reference drop on our dentry lock. Right now this
3890 * happens in the downconvert thread, but we could choose to simplify the
3891 * dlmglue API and push these off to the ocfs2_wq in the future.
3892 */
3893static void ocfs2_dentry_post_unlock(struct ocfs2_super *osb,
3894 struct ocfs2_lock_res *lockres)
3895{
3896 struct ocfs2_dentry_lock *dl = ocfs2_lock_res_dl(lockres);
3897 ocfs2_dentry_lock_put(osb, dl);
3898}
3899
3900/*
3901 * d_delete() matching dentries before the lock downconvert.
3902 *
3903 * At this point, any process waiting to destroy the
3904 * dentry_lock due to last ref count is stopped by the
3905 * OCFS2_LOCK_QUEUED flag.
3906 *
3907 * We have two potential problems
3908 *
3909 * 1) If we do the last reference drop on our dentry_lock (via dput)
3910 * we'll wind up in ocfs2_release_dentry_lock(), waiting on
3911 * the downconvert to finish. Instead we take an elevated
3912 * reference and push the drop until after we've completed our
3913 * unblock processing.
3914 *
3915 * 2) There might be another process with a final reference,
3916 * waiting on us to finish processing. If this is the case, we
3917 * detect it and exit out - there's no more dentries anyway.
3918 */
3919static int ocfs2_dentry_convert_worker(struct ocfs2_lock_res *lockres,
3920 int blocking)
3921{
3922 struct ocfs2_dentry_lock *dl = ocfs2_lock_res_dl(lockres);
3923 struct ocfs2_inode_info *oi = OCFS2_I(dl->dl_inode);
3924 struct dentry *dentry;
3925 unsigned long flags;
3926 int extra_ref = 0;
3927
3928 /*
3929 * This node is blocking another node from getting a read
3930 * lock. This happens when we've renamed within a
3931 * directory. We've forced the other nodes to d_delete(), but
3932 * we never actually dropped our lock because it's still
3933 * valid. The downconvert code will retain a PR for this node,
3934 * so there's no further work to do.
3935 */
3936 if (blocking == DLM_LOCK_PR)
3937 return UNBLOCK_CONTINUE;
3938
3939 /*
3940 * Mark this inode as potentially orphaned. The code in
3941 * ocfs2_delete_inode() will figure out whether it actually
3942 * needs to be freed or not.
3943 */
3944 spin_lock(&oi->ip_lock);
3945 oi->ip_flags |= OCFS2_INODE_MAYBE_ORPHANED;
3946 spin_unlock(&oi->ip_lock);
3947
3948 /*
3949 * Yuck. We need to make sure however that the check of
3950 * OCFS2_LOCK_FREEING and the extra reference are atomic with
3951 * respect to a reference decrement or the setting of that
3952 * flag.
3953 */
3954 spin_lock_irqsave(&lockres->l_lock, flags);
3955 spin_lock(&dentry_attach_lock);
3956 if (!(lockres->l_flags & OCFS2_LOCK_FREEING)
3957 && dl->dl_count) {
3958 dl->dl_count++;
3959 extra_ref = 1;
3960 }
3961 spin_unlock(&dentry_attach_lock);
3962 spin_unlock_irqrestore(&lockres->l_lock, flags);
3963
3964 mlog(0, "extra_ref = %d\n", extra_ref);
3965
3966 /*
3967 * We have a process waiting on us in ocfs2_dentry_iput(),
3968 * which means we can't have any more outstanding
3969 * aliases. There's no need to do any more work.
3970 */
3971 if (!extra_ref)
3972 return UNBLOCK_CONTINUE;
3973
3974 spin_lock(&dentry_attach_lock);
3975 while (1) {
3976 dentry = ocfs2_find_local_alias(dl->dl_inode,
3977 dl->dl_parent_blkno, 1);
3978 if (!dentry)
3979 break;
3980 spin_unlock(&dentry_attach_lock);
3981
3982 if (S_ISDIR(dl->dl_inode->i_mode))
3983 shrink_dcache_parent(dentry);
3984
3985 mlog(0, "d_delete(%pd);\n", dentry);
3986
3987 /*
3988 * The following dcache calls may do an
3989 * iput(). Normally we don't want that from the
3990 * downconverting thread, but in this case it's ok
3991 * because the requesting node already has an
3992 * exclusive lock on the inode, so it can't be queued
3993 * for a downconvert.
3994 */
3995 d_delete(dentry);
3996 dput(dentry);
3997
3998 spin_lock(&dentry_attach_lock);
3999 }
4000 spin_unlock(&dentry_attach_lock);
4001
4002 /*
4003 * If we are the last holder of this dentry lock, there is no
4004 * reason to downconvert so skip straight to the unlock.
4005 */
4006 if (dl->dl_count == 1)
4007 return UNBLOCK_STOP_POST;
4008
4009 return UNBLOCK_CONTINUE_POST;
4010}
4011
4012static int ocfs2_check_refcount_downconvert(struct ocfs2_lock_res *lockres,
4013 int new_level)
4014{
4015 struct ocfs2_refcount_tree *tree =
4016 ocfs2_lock_res_refcount_tree(lockres);
4017
4018 return ocfs2_ci_checkpointed(&tree->rf_ci, lockres, new_level);
4019}
4020
4021static int ocfs2_refcount_convert_worker(struct ocfs2_lock_res *lockres,
4022 int blocking)
4023{
4024 struct ocfs2_refcount_tree *tree =
4025 ocfs2_lock_res_refcount_tree(lockres);
4026
4027 ocfs2_metadata_cache_purge(&tree->rf_ci);
4028
4029 return UNBLOCK_CONTINUE;
4030}
4031
4032static void ocfs2_set_qinfo_lvb(struct ocfs2_lock_res *lockres)
4033{
4034 struct ocfs2_qinfo_lvb *lvb;
4035 struct ocfs2_mem_dqinfo *oinfo = ocfs2_lock_res_qinfo(lockres);
4036 struct mem_dqinfo *info = sb_dqinfo(oinfo->dqi_gi.dqi_sb,
4037 oinfo->dqi_gi.dqi_type);
4038
4039 lvb = ocfs2_dlm_lvb(&lockres->l_lksb);
4040 lvb->lvb_version = OCFS2_QINFO_LVB_VERSION;
4041 lvb->lvb_bgrace = cpu_to_be32(info->dqi_bgrace);
4042 lvb->lvb_igrace = cpu_to_be32(info->dqi_igrace);
4043 lvb->lvb_syncms = cpu_to_be32(oinfo->dqi_syncms);
4044 lvb->lvb_blocks = cpu_to_be32(oinfo->dqi_gi.dqi_blocks);
4045 lvb->lvb_free_blk = cpu_to_be32(oinfo->dqi_gi.dqi_free_blk);
4046 lvb->lvb_free_entry = cpu_to_be32(oinfo->dqi_gi.dqi_free_entry);
4047}
4048
4049void ocfs2_qinfo_unlock(struct ocfs2_mem_dqinfo *oinfo, int ex)
4050{
4051 struct ocfs2_lock_res *lockres = &oinfo->dqi_gqlock;
4052 struct ocfs2_super *osb = OCFS2_SB(oinfo->dqi_gi.dqi_sb);
4053 int level = ex ? DLM_LOCK_EX : DLM_LOCK_PR;
4054
4055 if (!ocfs2_is_hard_readonly(osb) && !ocfs2_mount_local(osb))
4056 ocfs2_cluster_unlock(osb, lockres, level);
4057}
4058
4059static int ocfs2_refresh_qinfo(struct ocfs2_mem_dqinfo *oinfo)
4060{
4061 struct mem_dqinfo *info = sb_dqinfo(oinfo->dqi_gi.dqi_sb,
4062 oinfo->dqi_gi.dqi_type);
4063 struct ocfs2_lock_res *lockres = &oinfo->dqi_gqlock;
4064 struct ocfs2_qinfo_lvb *lvb = ocfs2_dlm_lvb(&lockres->l_lksb);
4065 struct buffer_head *bh = NULL;
4066 struct ocfs2_global_disk_dqinfo *gdinfo;
4067 int status = 0;
4068
4069 if (ocfs2_dlm_lvb_valid(&lockres->l_lksb) &&
4070 lvb->lvb_version == OCFS2_QINFO_LVB_VERSION) {
4071 info->dqi_bgrace = be32_to_cpu(lvb->lvb_bgrace);
4072 info->dqi_igrace = be32_to_cpu(lvb->lvb_igrace);
4073 oinfo->dqi_syncms = be32_to_cpu(lvb->lvb_syncms);
4074 oinfo->dqi_gi.dqi_blocks = be32_to_cpu(lvb->lvb_blocks);
4075 oinfo->dqi_gi.dqi_free_blk = be32_to_cpu(lvb->lvb_free_blk);
4076 oinfo->dqi_gi.dqi_free_entry =
4077 be32_to_cpu(lvb->lvb_free_entry);
4078 } else {
4079 status = ocfs2_read_quota_phys_block(oinfo->dqi_gqinode,
4080 oinfo->dqi_giblk, &bh);
4081 if (status) {
4082 mlog_errno(status);
4083 goto bail;
4084 }
4085 gdinfo = (struct ocfs2_global_disk_dqinfo *)
4086 (bh->b_data + OCFS2_GLOBAL_INFO_OFF);
4087 info->dqi_bgrace = le32_to_cpu(gdinfo->dqi_bgrace);
4088 info->dqi_igrace = le32_to_cpu(gdinfo->dqi_igrace);
4089 oinfo->dqi_syncms = le32_to_cpu(gdinfo->dqi_syncms);
4090 oinfo->dqi_gi.dqi_blocks = le32_to_cpu(gdinfo->dqi_blocks);
4091 oinfo->dqi_gi.dqi_free_blk = le32_to_cpu(gdinfo->dqi_free_blk);
4092 oinfo->dqi_gi.dqi_free_entry =
4093 le32_to_cpu(gdinfo->dqi_free_entry);
4094 brelse(bh);
4095 ocfs2_track_lock_refresh(lockres);
4096 }
4097
4098bail:
4099 return status;
4100}
4101
4102/* Lock quota info, this function expects at least shared lock on the quota file
4103 * so that we can safely refresh quota info from disk. */
4104int ocfs2_qinfo_lock(struct ocfs2_mem_dqinfo *oinfo, int ex)
4105{
4106 struct ocfs2_lock_res *lockres = &oinfo->dqi_gqlock;
4107 struct ocfs2_super *osb = OCFS2_SB(oinfo->dqi_gi.dqi_sb);
4108 int level = ex ? DLM_LOCK_EX : DLM_LOCK_PR;
4109 int status = 0;
4110
4111 /* On RO devices, locking really isn't needed... */
4112 if (ocfs2_is_hard_readonly(osb)) {
4113 if (ex)
4114 status = -EROFS;
4115 goto bail;
4116 }
4117 if (ocfs2_mount_local(osb))
4118 goto bail;
4119
4120 status = ocfs2_cluster_lock(osb, lockres, level, 0, 0);
4121 if (status < 0) {
4122 mlog_errno(status);
4123 goto bail;
4124 }
4125 if (!ocfs2_should_refresh_lock_res(lockres))
4126 goto bail;
4127 /* OK, we have the lock but we need to refresh the quota info */
4128 status = ocfs2_refresh_qinfo(oinfo);
4129 if (status)
4130 ocfs2_qinfo_unlock(oinfo, ex);
4131 ocfs2_complete_lock_res_refresh(lockres, status);
4132bail:
4133 return status;
4134}
4135
4136int ocfs2_refcount_lock(struct ocfs2_refcount_tree *ref_tree, int ex)
4137{
4138 int status;
4139 int level = ex ? DLM_LOCK_EX : DLM_LOCK_PR;
4140 struct ocfs2_lock_res *lockres = &ref_tree->rf_lockres;
4141 struct ocfs2_super *osb = lockres->l_priv;
4142
4143
4144 if (ocfs2_is_hard_readonly(osb))
4145 return -EROFS;
4146
4147 if (ocfs2_mount_local(osb))
4148 return 0;
4149
4150 status = ocfs2_cluster_lock(osb, lockres, level, 0, 0);
4151 if (status < 0)
4152 mlog_errno(status);
4153
4154 return status;
4155}
4156
4157void ocfs2_refcount_unlock(struct ocfs2_refcount_tree *ref_tree, int ex)
4158{
4159 int level = ex ? DLM_LOCK_EX : DLM_LOCK_PR;
4160 struct ocfs2_lock_res *lockres = &ref_tree->rf_lockres;
4161 struct ocfs2_super *osb = lockres->l_priv;
4162
4163 if (!ocfs2_mount_local(osb))
4164 ocfs2_cluster_unlock(osb, lockres, level);
4165}
4166
4167static void ocfs2_process_blocked_lock(struct ocfs2_super *osb,
4168 struct ocfs2_lock_res *lockres)
4169{
4170 int status;
4171 struct ocfs2_unblock_ctl ctl = {0, 0,};
4172 unsigned long flags;
4173
4174 /* Our reference to the lockres in this function can be
4175 * considered valid until we remove the OCFS2_LOCK_QUEUED
4176 * flag. */
4177
4178 BUG_ON(!lockres);
4179 BUG_ON(!lockres->l_ops);
4180
4181 mlog(ML_BASTS, "lockres %s blocked\n", lockres->l_name);
4182
4183 /* Detect whether a lock has been marked as going away while
4184 * the downconvert thread was processing other things. A lock can
4185 * still be marked with OCFS2_LOCK_FREEING after this check,
4186 * but short circuiting here will still save us some
4187 * performance. */
4188 spin_lock_irqsave(&lockres->l_lock, flags);
4189 if (lockres->l_flags & OCFS2_LOCK_FREEING)
4190 goto unqueue;
4191 spin_unlock_irqrestore(&lockres->l_lock, flags);
4192
4193 status = ocfs2_unblock_lock(osb, lockres, &ctl);
4194 if (status < 0)
4195 mlog_errno(status);
4196
4197 spin_lock_irqsave(&lockres->l_lock, flags);
4198unqueue:
4199 if (lockres->l_flags & OCFS2_LOCK_FREEING || !ctl.requeue) {
4200 lockres_clear_flags(lockres, OCFS2_LOCK_QUEUED);
4201 } else
4202 ocfs2_schedule_blocked_lock(osb, lockres);
4203
4204 mlog(ML_BASTS, "lockres %s, requeue = %s.\n", lockres->l_name,
4205 ctl.requeue ? "yes" : "no");
4206 spin_unlock_irqrestore(&lockres->l_lock, flags);
4207
4208 if (ctl.unblock_action != UNBLOCK_CONTINUE
4209 && lockres->l_ops->post_unlock)
4210 lockres->l_ops->post_unlock(osb, lockres);
4211}
4212
4213static void ocfs2_schedule_blocked_lock(struct ocfs2_super *osb,
4214 struct ocfs2_lock_res *lockres)
4215{
4216 unsigned long flags;
4217
4218 assert_spin_locked(&lockres->l_lock);
4219
4220 if (lockres->l_flags & OCFS2_LOCK_FREEING) {
4221 /* Do not schedule a lock for downconvert when it's on
4222 * the way to destruction - any nodes wanting access
4223 * to the resource will get it soon. */
4224 mlog(ML_BASTS, "lockres %s won't be scheduled: flags 0x%lx\n",
4225 lockres->l_name, lockres->l_flags);
4226 return;
4227 }
4228
4229 lockres_or_flags(lockres, OCFS2_LOCK_QUEUED);
4230
4231 spin_lock_irqsave(&osb->dc_task_lock, flags);
4232 if (list_empty(&lockres->l_blocked_list)) {
4233 list_add_tail(&lockres->l_blocked_list,
4234 &osb->blocked_lock_list);
4235 osb->blocked_lock_count++;
4236 }
4237 spin_unlock_irqrestore(&osb->dc_task_lock, flags);
4238}
4239
4240static void ocfs2_downconvert_thread_do_work(struct ocfs2_super *osb)
4241{
4242 unsigned long processed;
4243 unsigned long flags;
4244 struct ocfs2_lock_res *lockres;
4245
4246 spin_lock_irqsave(&osb->dc_task_lock, flags);
4247 /* grab this early so we know to try again if a state change and
4248 * wake happens part-way through our work */
4249 osb->dc_work_sequence = osb->dc_wake_sequence;
4250
4251 processed = osb->blocked_lock_count;
4252 /*
4253 * blocked lock processing in this loop might call iput which can
4254 * remove items off osb->blocked_lock_list. Downconvert up to
4255 * 'processed' number of locks, but stop short if we had some
4256 * removed in ocfs2_mark_lockres_freeing when downconverting.
4257 */
4258 while (processed && !list_empty(&osb->blocked_lock_list)) {
4259 lockres = list_entry(osb->blocked_lock_list.next,
4260 struct ocfs2_lock_res, l_blocked_list);
4261 list_del_init(&lockres->l_blocked_list);
4262 osb->blocked_lock_count--;
4263 spin_unlock_irqrestore(&osb->dc_task_lock, flags);
4264
4265 BUG_ON(!processed);
4266 processed--;
4267
4268 ocfs2_process_blocked_lock(osb, lockres);
4269
4270 spin_lock_irqsave(&osb->dc_task_lock, flags);
4271 }
4272 spin_unlock_irqrestore(&osb->dc_task_lock, flags);
4273}
4274
4275static int ocfs2_downconvert_thread_lists_empty(struct ocfs2_super *osb)
4276{
4277 int empty = 0;
4278 unsigned long flags;
4279
4280 spin_lock_irqsave(&osb->dc_task_lock, flags);
4281 if (list_empty(&osb->blocked_lock_list))
4282 empty = 1;
4283
4284 spin_unlock_irqrestore(&osb->dc_task_lock, flags);
4285 return empty;
4286}
4287
4288static int ocfs2_downconvert_thread_should_wake(struct ocfs2_super *osb)
4289{
4290 int should_wake = 0;
4291 unsigned long flags;
4292
4293 spin_lock_irqsave(&osb->dc_task_lock, flags);
4294 if (osb->dc_work_sequence != osb->dc_wake_sequence)
4295 should_wake = 1;
4296 spin_unlock_irqrestore(&osb->dc_task_lock, flags);
4297
4298 return should_wake;
4299}
4300
4301static int ocfs2_downconvert_thread(void *arg)
4302{
4303 int status = 0;
4304 struct ocfs2_super *osb = arg;
4305
4306 /* only quit once we've been asked to stop and there is no more
4307 * work available */
4308 while (!(kthread_should_stop() &&
4309 ocfs2_downconvert_thread_lists_empty(osb))) {
4310
4311 wait_event_interruptible(osb->dc_event,
4312 ocfs2_downconvert_thread_should_wake(osb) ||
4313 kthread_should_stop());
4314
4315 mlog(0, "downconvert_thread: awoken\n");
4316
4317 ocfs2_downconvert_thread_do_work(osb);
4318 }
4319
4320 osb->dc_task = NULL;
4321 return status;
4322}
4323
4324void ocfs2_wake_downconvert_thread(struct ocfs2_super *osb)
4325{
4326 unsigned long flags;
4327
4328 spin_lock_irqsave(&osb->dc_task_lock, flags);
4329 /* make sure the voting thread gets a swipe at whatever changes
4330 * the caller may have made to the voting state */
4331 osb->dc_wake_sequence++;
4332 spin_unlock_irqrestore(&osb->dc_task_lock, flags);
4333 wake_up(&osb->dc_event);
4334}
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
58#include "buffer_head_io.h"
59
60struct ocfs2_mask_waiter {
61 struct list_head mw_item;
62 int mw_status;
63 struct completion mw_complete;
64 unsigned long mw_mask;
65 unsigned long mw_goal;
66#ifdef CONFIG_OCFS2_FS_STATS
67 ktime_t mw_lock_start;
68#endif
69};
70
71static struct ocfs2_super *ocfs2_get_dentry_osb(struct ocfs2_lock_res *lockres);
72static struct ocfs2_super *ocfs2_get_inode_osb(struct ocfs2_lock_res *lockres);
73static struct ocfs2_super *ocfs2_get_file_osb(struct ocfs2_lock_res *lockres);
74static struct ocfs2_super *ocfs2_get_qinfo_osb(struct ocfs2_lock_res *lockres);
75
76/*
77 * Return value from ->downconvert_worker functions.
78 *
79 * These control the precise actions of ocfs2_unblock_lock()
80 * and ocfs2_process_blocked_lock()
81 *
82 */
83enum ocfs2_unblock_action {
84 UNBLOCK_CONTINUE = 0, /* Continue downconvert */
85 UNBLOCK_CONTINUE_POST = 1, /* Continue downconvert, fire
86 * ->post_unlock callback */
87 UNBLOCK_STOP_POST = 2, /* Do not downconvert, fire
88 * ->post_unlock() callback. */
89};
90
91struct ocfs2_unblock_ctl {
92 int requeue;
93 enum ocfs2_unblock_action unblock_action;
94};
95
96/* Lockdep class keys */
97struct lock_class_key lockdep_keys[OCFS2_NUM_LOCK_TYPES];
98
99static int ocfs2_check_meta_downconvert(struct ocfs2_lock_res *lockres,
100 int new_level);
101static void ocfs2_set_meta_lvb(struct ocfs2_lock_res *lockres);
102
103static int ocfs2_data_convert_worker(struct ocfs2_lock_res *lockres,
104 int blocking);
105
106static int ocfs2_dentry_convert_worker(struct ocfs2_lock_res *lockres,
107 int blocking);
108
109static void ocfs2_dentry_post_unlock(struct ocfs2_super *osb,
110 struct ocfs2_lock_res *lockres);
111
112static void ocfs2_set_qinfo_lvb(struct ocfs2_lock_res *lockres);
113
114static int ocfs2_check_refcount_downconvert(struct ocfs2_lock_res *lockres,
115 int new_level);
116static int ocfs2_refcount_convert_worker(struct ocfs2_lock_res *lockres,
117 int blocking);
118
119#define mlog_meta_lvb(__level, __lockres) ocfs2_dump_meta_lvb_info(__level, __PRETTY_FUNCTION__, __LINE__, __lockres)
120
121/* This aids in debugging situations where a bad LVB might be involved. */
122static void ocfs2_dump_meta_lvb_info(u64 level,
123 const char *function,
124 unsigned int line,
125 struct ocfs2_lock_res *lockres)
126{
127 struct ocfs2_meta_lvb *lvb = ocfs2_dlm_lvb(&lockres->l_lksb);
128
129 mlog(level, "LVB information for %s (called from %s:%u):\n",
130 lockres->l_name, function, line);
131 mlog(level, "version: %u, clusters: %u, generation: 0x%x\n",
132 lvb->lvb_version, be32_to_cpu(lvb->lvb_iclusters),
133 be32_to_cpu(lvb->lvb_igeneration));
134 mlog(level, "size: %llu, uid %u, gid %u, mode 0x%x\n",
135 (unsigned long long)be64_to_cpu(lvb->lvb_isize),
136 be32_to_cpu(lvb->lvb_iuid), be32_to_cpu(lvb->lvb_igid),
137 be16_to_cpu(lvb->lvb_imode));
138 mlog(level, "nlink %u, atime_packed 0x%llx, ctime_packed 0x%llx, "
139 "mtime_packed 0x%llx iattr 0x%x\n", be16_to_cpu(lvb->lvb_inlink),
140 (long long)be64_to_cpu(lvb->lvb_iatime_packed),
141 (long long)be64_to_cpu(lvb->lvb_ictime_packed),
142 (long long)be64_to_cpu(lvb->lvb_imtime_packed),
143 be32_to_cpu(lvb->lvb_iattr));
144}
145
146
147/*
148 * OCFS2 Lock Resource Operations
149 *
150 * These fine tune the behavior of the generic dlmglue locking infrastructure.
151 *
152 * The most basic of lock types can point ->l_priv to their respective
153 * struct ocfs2_super and allow the default actions to manage things.
154 *
155 * Right now, each lock type also needs to implement an init function,
156 * and trivial lock/unlock wrappers. ocfs2_simple_drop_lockres()
157 * should be called when the lock is no longer needed (i.e., object
158 * destruction time).
159 */
160struct ocfs2_lock_res_ops {
161 /*
162 * Translate an ocfs2_lock_res * into an ocfs2_super *. Define
163 * this callback if ->l_priv is not an ocfs2_super pointer
164 */
165 struct ocfs2_super * (*get_osb)(struct ocfs2_lock_res *);
166
167 /*
168 * Optionally called in the downconvert thread after a
169 * successful downconvert. The lockres will not be referenced
170 * after this callback is called, so it is safe to free
171 * memory, etc.
172 *
173 * The exact semantics of when this is called are controlled
174 * by ->downconvert_worker()
175 */
176 void (*post_unlock)(struct ocfs2_super *, struct ocfs2_lock_res *);
177
178 /*
179 * Allow a lock type to add checks to determine whether it is
180 * safe to downconvert a lock. Return 0 to re-queue the
181 * downconvert at a later time, nonzero to continue.
182 *
183 * For most locks, the default checks that there are no
184 * incompatible holders are sufficient.
185 *
186 * Called with the lockres spinlock held.
187 */
188 int (*check_downconvert)(struct ocfs2_lock_res *, int);
189
190 /*
191 * Allows a lock type to populate the lock value block. This
192 * is called on downconvert, and when we drop a lock.
193 *
194 * Locks that want to use this should set LOCK_TYPE_USES_LVB
195 * in the flags field.
196 *
197 * Called with the lockres spinlock held.
198 */
199 void (*set_lvb)(struct ocfs2_lock_res *);
200
201 /*
202 * Called from the downconvert thread when it is determined
203 * that a lock will be downconverted. This is called without
204 * any locks held so the function can do work that might
205 * schedule (syncing out data, etc).
206 *
207 * This should return any one of the ocfs2_unblock_action
208 * values, depending on what it wants the thread to do.
209 */
210 int (*downconvert_worker)(struct ocfs2_lock_res *, int);
211
212 /*
213 * LOCK_TYPE_* flags which describe the specific requirements
214 * of a lock type. Descriptions of each individual flag follow.
215 */
216 int flags;
217};
218
219/*
220 * Some locks want to "refresh" potentially stale data when a
221 * meaningful (PRMODE or EXMODE) lock level is first obtained. If this
222 * flag is set, the OCFS2_LOCK_NEEDS_REFRESH flag will be set on the
223 * individual lockres l_flags member from the ast function. It is
224 * expected that the locking wrapper will clear the
225 * OCFS2_LOCK_NEEDS_REFRESH flag when done.
226 */
227#define LOCK_TYPE_REQUIRES_REFRESH 0x1
228
229/*
230 * Indicate that a lock type makes use of the lock value block. The
231 * ->set_lvb lock type callback must be defined.
232 */
233#define LOCK_TYPE_USES_LVB 0x2
234
235static struct ocfs2_lock_res_ops ocfs2_inode_rw_lops = {
236 .get_osb = ocfs2_get_inode_osb,
237 .flags = 0,
238};
239
240static struct ocfs2_lock_res_ops ocfs2_inode_inode_lops = {
241 .get_osb = ocfs2_get_inode_osb,
242 .check_downconvert = ocfs2_check_meta_downconvert,
243 .set_lvb = ocfs2_set_meta_lvb,
244 .downconvert_worker = ocfs2_data_convert_worker,
245 .flags = LOCK_TYPE_REQUIRES_REFRESH|LOCK_TYPE_USES_LVB,
246};
247
248static struct ocfs2_lock_res_ops ocfs2_super_lops = {
249 .flags = LOCK_TYPE_REQUIRES_REFRESH,
250};
251
252static struct ocfs2_lock_res_ops ocfs2_rename_lops = {
253 .flags = 0,
254};
255
256static struct ocfs2_lock_res_ops ocfs2_nfs_sync_lops = {
257 .flags = 0,
258};
259
260static struct ocfs2_lock_res_ops ocfs2_orphan_scan_lops = {
261 .flags = LOCK_TYPE_REQUIRES_REFRESH|LOCK_TYPE_USES_LVB,
262};
263
264static struct ocfs2_lock_res_ops ocfs2_dentry_lops = {
265 .get_osb = ocfs2_get_dentry_osb,
266 .post_unlock = ocfs2_dentry_post_unlock,
267 .downconvert_worker = ocfs2_dentry_convert_worker,
268 .flags = 0,
269};
270
271static struct ocfs2_lock_res_ops ocfs2_inode_open_lops = {
272 .get_osb = ocfs2_get_inode_osb,
273 .flags = 0,
274};
275
276static struct ocfs2_lock_res_ops ocfs2_flock_lops = {
277 .get_osb = ocfs2_get_file_osb,
278 .flags = 0,
279};
280
281static struct ocfs2_lock_res_ops ocfs2_qinfo_lops = {
282 .set_lvb = ocfs2_set_qinfo_lvb,
283 .get_osb = ocfs2_get_qinfo_osb,
284 .flags = LOCK_TYPE_REQUIRES_REFRESH | LOCK_TYPE_USES_LVB,
285};
286
287static struct ocfs2_lock_res_ops ocfs2_refcount_block_lops = {
288 .check_downconvert = ocfs2_check_refcount_downconvert,
289 .downconvert_worker = ocfs2_refcount_convert_worker,
290 .flags = 0,
291};
292
293static inline int ocfs2_is_inode_lock(struct ocfs2_lock_res *lockres)
294{
295 return lockres->l_type == OCFS2_LOCK_TYPE_META ||
296 lockres->l_type == OCFS2_LOCK_TYPE_RW ||
297 lockres->l_type == OCFS2_LOCK_TYPE_OPEN;
298}
299
300static inline struct ocfs2_lock_res *ocfs2_lksb_to_lock_res(struct ocfs2_dlm_lksb *lksb)
301{
302 return container_of(lksb, struct ocfs2_lock_res, l_lksb);
303}
304
305static inline struct inode *ocfs2_lock_res_inode(struct ocfs2_lock_res *lockres)
306{
307 BUG_ON(!ocfs2_is_inode_lock(lockres));
308
309 return (struct inode *) lockres->l_priv;
310}
311
312static inline struct ocfs2_dentry_lock *ocfs2_lock_res_dl(struct ocfs2_lock_res *lockres)
313{
314 BUG_ON(lockres->l_type != OCFS2_LOCK_TYPE_DENTRY);
315
316 return (struct ocfs2_dentry_lock *)lockres->l_priv;
317}
318
319static inline struct ocfs2_mem_dqinfo *ocfs2_lock_res_qinfo(struct ocfs2_lock_res *lockres)
320{
321 BUG_ON(lockres->l_type != OCFS2_LOCK_TYPE_QINFO);
322
323 return (struct ocfs2_mem_dqinfo *)lockres->l_priv;
324}
325
326static inline struct ocfs2_refcount_tree *
327ocfs2_lock_res_refcount_tree(struct ocfs2_lock_res *res)
328{
329 return container_of(res, struct ocfs2_refcount_tree, rf_lockres);
330}
331
332static inline struct ocfs2_super *ocfs2_get_lockres_osb(struct ocfs2_lock_res *lockres)
333{
334 if (lockres->l_ops->get_osb)
335 return lockres->l_ops->get_osb(lockres);
336
337 return (struct ocfs2_super *)lockres->l_priv;
338}
339
340static int ocfs2_lock_create(struct ocfs2_super *osb,
341 struct ocfs2_lock_res *lockres,
342 int level,
343 u32 dlm_flags);
344static inline int ocfs2_may_continue_on_blocked_lock(struct ocfs2_lock_res *lockres,
345 int wanted);
346static void __ocfs2_cluster_unlock(struct ocfs2_super *osb,
347 struct ocfs2_lock_res *lockres,
348 int level, unsigned long caller_ip);
349static inline void ocfs2_cluster_unlock(struct ocfs2_super *osb,
350 struct ocfs2_lock_res *lockres,
351 int level)
352{
353 __ocfs2_cluster_unlock(osb, lockres, level, _RET_IP_);
354}
355
356static inline void ocfs2_generic_handle_downconvert_action(struct ocfs2_lock_res *lockres);
357static inline void ocfs2_generic_handle_convert_action(struct ocfs2_lock_res *lockres);
358static inline void ocfs2_generic_handle_attach_action(struct ocfs2_lock_res *lockres);
359static int ocfs2_generic_handle_bast(struct ocfs2_lock_res *lockres, int level);
360static void ocfs2_schedule_blocked_lock(struct ocfs2_super *osb,
361 struct ocfs2_lock_res *lockres);
362static inline void ocfs2_recover_from_dlm_error(struct ocfs2_lock_res *lockres,
363 int convert);
364#define ocfs2_log_dlm_error(_func, _err, _lockres) do { \
365 if ((_lockres)->l_type != OCFS2_LOCK_TYPE_DENTRY) \
366 mlog(ML_ERROR, "DLM error %d while calling %s on resource %s\n", \
367 _err, _func, _lockres->l_name); \
368 else \
369 mlog(ML_ERROR, "DLM error %d while calling %s on resource %.*s%08x\n", \
370 _err, _func, OCFS2_DENTRY_LOCK_INO_START - 1, (_lockres)->l_name, \
371 (unsigned int)ocfs2_get_dentry_lock_ino(_lockres)); \
372} while (0)
373static int ocfs2_downconvert_thread(void *arg);
374static void ocfs2_downconvert_on_unlock(struct ocfs2_super *osb,
375 struct ocfs2_lock_res *lockres);
376static int ocfs2_inode_lock_update(struct inode *inode,
377 struct buffer_head **bh);
378static void ocfs2_drop_osb_locks(struct ocfs2_super *osb);
379static inline int ocfs2_highest_compat_lock_level(int level);
380static unsigned int ocfs2_prepare_downconvert(struct ocfs2_lock_res *lockres,
381 int new_level);
382static int ocfs2_downconvert_lock(struct ocfs2_super *osb,
383 struct ocfs2_lock_res *lockres,
384 int new_level,
385 int lvb,
386 unsigned int generation);
387static int ocfs2_prepare_cancel_convert(struct ocfs2_super *osb,
388 struct ocfs2_lock_res *lockres);
389static int ocfs2_cancel_convert(struct ocfs2_super *osb,
390 struct ocfs2_lock_res *lockres);
391
392
393static void ocfs2_build_lock_name(enum ocfs2_lock_type type,
394 u64 blkno,
395 u32 generation,
396 char *name)
397{
398 int len;
399
400 BUG_ON(type >= OCFS2_NUM_LOCK_TYPES);
401
402 len = snprintf(name, OCFS2_LOCK_ID_MAX_LEN, "%c%s%016llx%08x",
403 ocfs2_lock_type_char(type), OCFS2_LOCK_ID_PAD,
404 (long long)blkno, generation);
405
406 BUG_ON(len != (OCFS2_LOCK_ID_MAX_LEN - 1));
407
408 mlog(0, "built lock resource with name: %s\n", name);
409}
410
411static DEFINE_SPINLOCK(ocfs2_dlm_tracking_lock);
412
413static void ocfs2_add_lockres_tracking(struct ocfs2_lock_res *res,
414 struct ocfs2_dlm_debug *dlm_debug)
415{
416 mlog(0, "Add tracking for lockres %s\n", res->l_name);
417
418 spin_lock(&ocfs2_dlm_tracking_lock);
419 list_add(&res->l_debug_list, &dlm_debug->d_lockres_tracking);
420 spin_unlock(&ocfs2_dlm_tracking_lock);
421}
422
423static void ocfs2_remove_lockres_tracking(struct ocfs2_lock_res *res)
424{
425 spin_lock(&ocfs2_dlm_tracking_lock);
426 if (!list_empty(&res->l_debug_list))
427 list_del_init(&res->l_debug_list);
428 spin_unlock(&ocfs2_dlm_tracking_lock);
429}
430
431#ifdef CONFIG_OCFS2_FS_STATS
432static void ocfs2_init_lock_stats(struct ocfs2_lock_res *res)
433{
434 res->l_lock_refresh = 0;
435 memset(&res->l_lock_prmode, 0, sizeof(struct ocfs2_lock_stats));
436 memset(&res->l_lock_exmode, 0, sizeof(struct ocfs2_lock_stats));
437}
438
439static void ocfs2_update_lock_stats(struct ocfs2_lock_res *res, int level,
440 struct ocfs2_mask_waiter *mw, int ret)
441{
442 u32 usec;
443 ktime_t kt;
444 struct ocfs2_lock_stats *stats;
445
446 if (level == LKM_PRMODE)
447 stats = &res->l_lock_prmode;
448 else if (level == LKM_EXMODE)
449 stats = &res->l_lock_exmode;
450 else
451 return;
452
453 kt = ktime_sub(ktime_get(), mw->mw_lock_start);
454 usec = ktime_to_us(kt);
455
456 stats->ls_gets++;
457 stats->ls_total += ktime_to_ns(kt);
458 /* overflow */
459 if (unlikely(stats->ls_gets) == 0) {
460 stats->ls_gets++;
461 stats->ls_total = ktime_to_ns(kt);
462 }
463
464 if (stats->ls_max < usec)
465 stats->ls_max = usec;
466
467 if (ret)
468 stats->ls_fail++;
469}
470
471static inline void ocfs2_track_lock_refresh(struct ocfs2_lock_res *lockres)
472{
473 lockres->l_lock_refresh++;
474}
475
476static inline void ocfs2_init_start_time(struct ocfs2_mask_waiter *mw)
477{
478 mw->mw_lock_start = ktime_get();
479}
480#else
481static inline void ocfs2_init_lock_stats(struct ocfs2_lock_res *res)
482{
483}
484static inline void ocfs2_update_lock_stats(struct ocfs2_lock_res *res,
485 int level, struct ocfs2_mask_waiter *mw, int ret)
486{
487}
488static inline void ocfs2_track_lock_refresh(struct ocfs2_lock_res *lockres)
489{
490}
491static inline void ocfs2_init_start_time(struct ocfs2_mask_waiter *mw)
492{
493}
494#endif
495
496static void ocfs2_lock_res_init_common(struct ocfs2_super *osb,
497 struct ocfs2_lock_res *res,
498 enum ocfs2_lock_type type,
499 struct ocfs2_lock_res_ops *ops,
500 void *priv)
501{
502 res->l_type = type;
503 res->l_ops = ops;
504 res->l_priv = priv;
505
506 res->l_level = DLM_LOCK_IV;
507 res->l_requested = DLM_LOCK_IV;
508 res->l_blocking = DLM_LOCK_IV;
509 res->l_action = OCFS2_AST_INVALID;
510 res->l_unlock_action = OCFS2_UNLOCK_INVALID;
511
512 res->l_flags = OCFS2_LOCK_INITIALIZED;
513
514 ocfs2_add_lockres_tracking(res, osb->osb_dlm_debug);
515
516 ocfs2_init_lock_stats(res);
517#ifdef CONFIG_DEBUG_LOCK_ALLOC
518 if (type != OCFS2_LOCK_TYPE_OPEN)
519 lockdep_init_map(&res->l_lockdep_map, ocfs2_lock_type_strings[type],
520 &lockdep_keys[type], 0);
521 else
522 res->l_lockdep_map.key = NULL;
523#endif
524}
525
526void ocfs2_lock_res_init_once(struct ocfs2_lock_res *res)
527{
528 /* This also clears out the lock status block */
529 memset(res, 0, sizeof(struct ocfs2_lock_res));
530 spin_lock_init(&res->l_lock);
531 init_waitqueue_head(&res->l_event);
532 INIT_LIST_HEAD(&res->l_blocked_list);
533 INIT_LIST_HEAD(&res->l_mask_waiters);
534}
535
536void ocfs2_inode_lock_res_init(struct ocfs2_lock_res *res,
537 enum ocfs2_lock_type type,
538 unsigned int generation,
539 struct inode *inode)
540{
541 struct ocfs2_lock_res_ops *ops;
542
543 switch(type) {
544 case OCFS2_LOCK_TYPE_RW:
545 ops = &ocfs2_inode_rw_lops;
546 break;
547 case OCFS2_LOCK_TYPE_META:
548 ops = &ocfs2_inode_inode_lops;
549 break;
550 case OCFS2_LOCK_TYPE_OPEN:
551 ops = &ocfs2_inode_open_lops;
552 break;
553 default:
554 mlog_bug_on_msg(1, "type: %d\n", type);
555 ops = NULL; /* thanks, gcc */
556 break;
557 };
558
559 ocfs2_build_lock_name(type, OCFS2_I(inode)->ip_blkno,
560 generation, res->l_name);
561 ocfs2_lock_res_init_common(OCFS2_SB(inode->i_sb), res, type, ops, inode);
562}
563
564static struct ocfs2_super *ocfs2_get_inode_osb(struct ocfs2_lock_res *lockres)
565{
566 struct inode *inode = ocfs2_lock_res_inode(lockres);
567
568 return OCFS2_SB(inode->i_sb);
569}
570
571static struct ocfs2_super *ocfs2_get_qinfo_osb(struct ocfs2_lock_res *lockres)
572{
573 struct ocfs2_mem_dqinfo *info = lockres->l_priv;
574
575 return OCFS2_SB(info->dqi_gi.dqi_sb);
576}
577
578static struct ocfs2_super *ocfs2_get_file_osb(struct ocfs2_lock_res *lockres)
579{
580 struct ocfs2_file_private *fp = lockres->l_priv;
581
582 return OCFS2_SB(fp->fp_file->f_mapping->host->i_sb);
583}
584
585static __u64 ocfs2_get_dentry_lock_ino(struct ocfs2_lock_res *lockres)
586{
587 __be64 inode_blkno_be;
588
589 memcpy(&inode_blkno_be, &lockres->l_name[OCFS2_DENTRY_LOCK_INO_START],
590 sizeof(__be64));
591
592 return be64_to_cpu(inode_blkno_be);
593}
594
595static struct ocfs2_super *ocfs2_get_dentry_osb(struct ocfs2_lock_res *lockres)
596{
597 struct ocfs2_dentry_lock *dl = lockres->l_priv;
598
599 return OCFS2_SB(dl->dl_inode->i_sb);
600}
601
602void ocfs2_dentry_lock_res_init(struct ocfs2_dentry_lock *dl,
603 u64 parent, struct inode *inode)
604{
605 int len;
606 u64 inode_blkno = OCFS2_I(inode)->ip_blkno;
607 __be64 inode_blkno_be = cpu_to_be64(inode_blkno);
608 struct ocfs2_lock_res *lockres = &dl->dl_lockres;
609
610 ocfs2_lock_res_init_once(lockres);
611
612 /*
613 * Unfortunately, the standard lock naming scheme won't work
614 * here because we have two 16 byte values to use. Instead,
615 * we'll stuff the inode number as a binary value. We still
616 * want error prints to show something without garbling the
617 * display, so drop a null byte in there before the inode
618 * number. A future version of OCFS2 will likely use all
619 * binary lock names. The stringified names have been a
620 * tremendous aid in debugging, but now that the debugfs
621 * interface exists, we can mangle things there if need be.
622 *
623 * NOTE: We also drop the standard "pad" value (the total lock
624 * name size stays the same though - the last part is all
625 * zeros due to the memset in ocfs2_lock_res_init_once()
626 */
627 len = snprintf(lockres->l_name, OCFS2_DENTRY_LOCK_INO_START,
628 "%c%016llx",
629 ocfs2_lock_type_char(OCFS2_LOCK_TYPE_DENTRY),
630 (long long)parent);
631
632 BUG_ON(len != (OCFS2_DENTRY_LOCK_INO_START - 1));
633
634 memcpy(&lockres->l_name[OCFS2_DENTRY_LOCK_INO_START], &inode_blkno_be,
635 sizeof(__be64));
636
637 ocfs2_lock_res_init_common(OCFS2_SB(inode->i_sb), lockres,
638 OCFS2_LOCK_TYPE_DENTRY, &ocfs2_dentry_lops,
639 dl);
640}
641
642static void ocfs2_super_lock_res_init(struct ocfs2_lock_res *res,
643 struct ocfs2_super *osb)
644{
645 /* Superblock lockres doesn't come from a slab so we call init
646 * once on it manually. */
647 ocfs2_lock_res_init_once(res);
648 ocfs2_build_lock_name(OCFS2_LOCK_TYPE_SUPER, OCFS2_SUPER_BLOCK_BLKNO,
649 0, res->l_name);
650 ocfs2_lock_res_init_common(osb, res, OCFS2_LOCK_TYPE_SUPER,
651 &ocfs2_super_lops, osb);
652}
653
654static void ocfs2_rename_lock_res_init(struct ocfs2_lock_res *res,
655 struct ocfs2_super *osb)
656{
657 /* Rename lockres doesn't come from a slab so we call init
658 * once on it manually. */
659 ocfs2_lock_res_init_once(res);
660 ocfs2_build_lock_name(OCFS2_LOCK_TYPE_RENAME, 0, 0, res->l_name);
661 ocfs2_lock_res_init_common(osb, res, OCFS2_LOCK_TYPE_RENAME,
662 &ocfs2_rename_lops, osb);
663}
664
665static void ocfs2_nfs_sync_lock_res_init(struct ocfs2_lock_res *res,
666 struct ocfs2_super *osb)
667{
668 /* nfs_sync lockres doesn't come from a slab so we call init
669 * once on it manually. */
670 ocfs2_lock_res_init_once(res);
671 ocfs2_build_lock_name(OCFS2_LOCK_TYPE_NFS_SYNC, 0, 0, res->l_name);
672 ocfs2_lock_res_init_common(osb, res, OCFS2_LOCK_TYPE_NFS_SYNC,
673 &ocfs2_nfs_sync_lops, osb);
674}
675
676static void ocfs2_orphan_scan_lock_res_init(struct ocfs2_lock_res *res,
677 struct ocfs2_super *osb)
678{
679 ocfs2_lock_res_init_once(res);
680 ocfs2_build_lock_name(OCFS2_LOCK_TYPE_ORPHAN_SCAN, 0, 0, res->l_name);
681 ocfs2_lock_res_init_common(osb, res, OCFS2_LOCK_TYPE_ORPHAN_SCAN,
682 &ocfs2_orphan_scan_lops, osb);
683}
684
685void ocfs2_file_lock_res_init(struct ocfs2_lock_res *lockres,
686 struct ocfs2_file_private *fp)
687{
688 struct inode *inode = fp->fp_file->f_mapping->host;
689 struct ocfs2_inode_info *oi = OCFS2_I(inode);
690
691 ocfs2_lock_res_init_once(lockres);
692 ocfs2_build_lock_name(OCFS2_LOCK_TYPE_FLOCK, oi->ip_blkno,
693 inode->i_generation, lockres->l_name);
694 ocfs2_lock_res_init_common(OCFS2_SB(inode->i_sb), lockres,
695 OCFS2_LOCK_TYPE_FLOCK, &ocfs2_flock_lops,
696 fp);
697 lockres->l_flags |= OCFS2_LOCK_NOCACHE;
698}
699
700void ocfs2_qinfo_lock_res_init(struct ocfs2_lock_res *lockres,
701 struct ocfs2_mem_dqinfo *info)
702{
703 ocfs2_lock_res_init_once(lockres);
704 ocfs2_build_lock_name(OCFS2_LOCK_TYPE_QINFO, info->dqi_gi.dqi_type,
705 0, lockres->l_name);
706 ocfs2_lock_res_init_common(OCFS2_SB(info->dqi_gi.dqi_sb), lockres,
707 OCFS2_LOCK_TYPE_QINFO, &ocfs2_qinfo_lops,
708 info);
709}
710
711void ocfs2_refcount_lock_res_init(struct ocfs2_lock_res *lockres,
712 struct ocfs2_super *osb, u64 ref_blkno,
713 unsigned int generation)
714{
715 ocfs2_lock_res_init_once(lockres);
716 ocfs2_build_lock_name(OCFS2_LOCK_TYPE_REFCOUNT, ref_blkno,
717 generation, lockres->l_name);
718 ocfs2_lock_res_init_common(osb, lockres, OCFS2_LOCK_TYPE_REFCOUNT,
719 &ocfs2_refcount_block_lops, osb);
720}
721
722void ocfs2_lock_res_free(struct ocfs2_lock_res *res)
723{
724 if (!(res->l_flags & OCFS2_LOCK_INITIALIZED))
725 return;
726
727 ocfs2_remove_lockres_tracking(res);
728
729 mlog_bug_on_msg(!list_empty(&res->l_blocked_list),
730 "Lockres %s is on the blocked list\n",
731 res->l_name);
732 mlog_bug_on_msg(!list_empty(&res->l_mask_waiters),
733 "Lockres %s has mask waiters pending\n",
734 res->l_name);
735 mlog_bug_on_msg(spin_is_locked(&res->l_lock),
736 "Lockres %s is locked\n",
737 res->l_name);
738 mlog_bug_on_msg(res->l_ro_holders,
739 "Lockres %s has %u ro holders\n",
740 res->l_name, res->l_ro_holders);
741 mlog_bug_on_msg(res->l_ex_holders,
742 "Lockres %s has %u ex holders\n",
743 res->l_name, res->l_ex_holders);
744
745 /* Need to clear out the lock status block for the dlm */
746 memset(&res->l_lksb, 0, sizeof(res->l_lksb));
747
748 res->l_flags = 0UL;
749}
750
751static inline void ocfs2_inc_holders(struct ocfs2_lock_res *lockres,
752 int level)
753{
754 BUG_ON(!lockres);
755
756 switch(level) {
757 case DLM_LOCK_EX:
758 lockres->l_ex_holders++;
759 break;
760 case DLM_LOCK_PR:
761 lockres->l_ro_holders++;
762 break;
763 default:
764 BUG();
765 }
766}
767
768static inline void ocfs2_dec_holders(struct ocfs2_lock_res *lockres,
769 int level)
770{
771 BUG_ON(!lockres);
772
773 switch(level) {
774 case DLM_LOCK_EX:
775 BUG_ON(!lockres->l_ex_holders);
776 lockres->l_ex_holders--;
777 break;
778 case DLM_LOCK_PR:
779 BUG_ON(!lockres->l_ro_holders);
780 lockres->l_ro_holders--;
781 break;
782 default:
783 BUG();
784 }
785}
786
787/* WARNING: This function lives in a world where the only three lock
788 * levels are EX, PR, and NL. It *will* have to be adjusted when more
789 * lock types are added. */
790static inline int ocfs2_highest_compat_lock_level(int level)
791{
792 int new_level = DLM_LOCK_EX;
793
794 if (level == DLM_LOCK_EX)
795 new_level = DLM_LOCK_NL;
796 else if (level == DLM_LOCK_PR)
797 new_level = DLM_LOCK_PR;
798 return new_level;
799}
800
801static void lockres_set_flags(struct ocfs2_lock_res *lockres,
802 unsigned long newflags)
803{
804 struct ocfs2_mask_waiter *mw, *tmp;
805
806 assert_spin_locked(&lockres->l_lock);
807
808 lockres->l_flags = newflags;
809
810 list_for_each_entry_safe(mw, tmp, &lockres->l_mask_waiters, mw_item) {
811 if ((lockres->l_flags & mw->mw_mask) != mw->mw_goal)
812 continue;
813
814 list_del_init(&mw->mw_item);
815 mw->mw_status = 0;
816 complete(&mw->mw_complete);
817 }
818}
819static void lockres_or_flags(struct ocfs2_lock_res *lockres, unsigned long or)
820{
821 lockres_set_flags(lockres, lockres->l_flags | or);
822}
823static void lockres_clear_flags(struct ocfs2_lock_res *lockres,
824 unsigned long clear)
825{
826 lockres_set_flags(lockres, lockres->l_flags & ~clear);
827}
828
829static inline void ocfs2_generic_handle_downconvert_action(struct ocfs2_lock_res *lockres)
830{
831 BUG_ON(!(lockres->l_flags & OCFS2_LOCK_BUSY));
832 BUG_ON(!(lockres->l_flags & OCFS2_LOCK_ATTACHED));
833 BUG_ON(!(lockres->l_flags & OCFS2_LOCK_BLOCKED));
834 BUG_ON(lockres->l_blocking <= DLM_LOCK_NL);
835
836 lockres->l_level = lockres->l_requested;
837 if (lockres->l_level <=
838 ocfs2_highest_compat_lock_level(lockres->l_blocking)) {
839 lockres->l_blocking = DLM_LOCK_NL;
840 lockres_clear_flags(lockres, OCFS2_LOCK_BLOCKED);
841 }
842 lockres_clear_flags(lockres, OCFS2_LOCK_BUSY);
843}
844
845static inline void ocfs2_generic_handle_convert_action(struct ocfs2_lock_res *lockres)
846{
847 BUG_ON(!(lockres->l_flags & OCFS2_LOCK_BUSY));
848 BUG_ON(!(lockres->l_flags & OCFS2_LOCK_ATTACHED));
849
850 /* Convert from RO to EX doesn't really need anything as our
851 * information is already up to data. Convert from NL to
852 * *anything* however should mark ourselves as needing an
853 * update */
854 if (lockres->l_level == DLM_LOCK_NL &&
855 lockres->l_ops->flags & LOCK_TYPE_REQUIRES_REFRESH)
856 lockres_or_flags(lockres, OCFS2_LOCK_NEEDS_REFRESH);
857
858 lockres->l_level = lockres->l_requested;
859
860 /*
861 * We set the OCFS2_LOCK_UPCONVERT_FINISHING flag before clearing
862 * the OCFS2_LOCK_BUSY flag to prevent the dc thread from
863 * downconverting the lock before the upconvert has fully completed.
864 */
865 lockres_or_flags(lockres, OCFS2_LOCK_UPCONVERT_FINISHING);
866
867 lockres_clear_flags(lockres, OCFS2_LOCK_BUSY);
868}
869
870static inline void ocfs2_generic_handle_attach_action(struct ocfs2_lock_res *lockres)
871{
872 BUG_ON((!(lockres->l_flags & OCFS2_LOCK_BUSY)));
873 BUG_ON(lockres->l_flags & OCFS2_LOCK_ATTACHED);
874
875 if (lockres->l_requested > DLM_LOCK_NL &&
876 !(lockres->l_flags & OCFS2_LOCK_LOCAL) &&
877 lockres->l_ops->flags & LOCK_TYPE_REQUIRES_REFRESH)
878 lockres_or_flags(lockres, OCFS2_LOCK_NEEDS_REFRESH);
879
880 lockres->l_level = lockres->l_requested;
881 lockres_or_flags(lockres, OCFS2_LOCK_ATTACHED);
882 lockres_clear_flags(lockres, OCFS2_LOCK_BUSY);
883}
884
885static int ocfs2_generic_handle_bast(struct ocfs2_lock_res *lockres,
886 int level)
887{
888 int needs_downconvert = 0;
889
890 assert_spin_locked(&lockres->l_lock);
891
892 if (level > lockres->l_blocking) {
893 /* only schedule a downconvert if we haven't already scheduled
894 * one that goes low enough to satisfy the level we're
895 * blocking. this also catches the case where we get
896 * duplicate BASTs */
897 if (ocfs2_highest_compat_lock_level(level) <
898 ocfs2_highest_compat_lock_level(lockres->l_blocking))
899 needs_downconvert = 1;
900
901 lockres->l_blocking = level;
902 }
903
904 mlog(ML_BASTS, "lockres %s, block %d, level %d, l_block %d, dwn %d\n",
905 lockres->l_name, level, lockres->l_level, lockres->l_blocking,
906 needs_downconvert);
907
908 if (needs_downconvert)
909 lockres_or_flags(lockres, OCFS2_LOCK_BLOCKED);
910 mlog(0, "needs_downconvert = %d\n", needs_downconvert);
911 return needs_downconvert;
912}
913
914/*
915 * OCFS2_LOCK_PENDING and l_pending_gen.
916 *
917 * Why does OCFS2_LOCK_PENDING exist? To close a race between setting
918 * OCFS2_LOCK_BUSY and calling ocfs2_dlm_lock(). See ocfs2_unblock_lock()
919 * for more details on the race.
920 *
921 * OCFS2_LOCK_PENDING closes the race quite nicely. However, it introduces
922 * a race on itself. In o2dlm, we can get the ast before ocfs2_dlm_lock()
923 * returns. The ast clears OCFS2_LOCK_BUSY, and must therefore clear
924 * OCFS2_LOCK_PENDING at the same time. When ocfs2_dlm_lock() returns,
925 * the caller is going to try to clear PENDING again. If nothing else is
926 * happening, __lockres_clear_pending() sees PENDING is unset and does
927 * nothing.
928 *
929 * But what if another path (eg downconvert thread) has just started a
930 * new locking action? The other path has re-set PENDING. Our path
931 * cannot clear PENDING, because that will re-open the original race
932 * window.
933 *
934 * [Example]
935 *
936 * ocfs2_meta_lock()
937 * ocfs2_cluster_lock()
938 * set BUSY
939 * set PENDING
940 * drop l_lock
941 * ocfs2_dlm_lock()
942 * ocfs2_locking_ast() ocfs2_downconvert_thread()
943 * clear PENDING ocfs2_unblock_lock()
944 * take_l_lock
945 * !BUSY
946 * ocfs2_prepare_downconvert()
947 * set BUSY
948 * set PENDING
949 * drop l_lock
950 * take l_lock
951 * clear PENDING
952 * drop l_lock
953 * <window>
954 * ocfs2_dlm_lock()
955 *
956 * So as you can see, we now have a window where l_lock is not held,
957 * PENDING is not set, and ocfs2_dlm_lock() has not been called.
958 *
959 * The core problem is that ocfs2_cluster_lock() has cleared the PENDING
960 * set by ocfs2_prepare_downconvert(). That wasn't nice.
961 *
962 * To solve this we introduce l_pending_gen. A call to
963 * lockres_clear_pending() will only do so when it is passed a generation
964 * number that matches the lockres. lockres_set_pending() will return the
965 * current generation number. When ocfs2_cluster_lock() goes to clear
966 * PENDING, it passes the generation it got from set_pending(). In our
967 * example above, the generation numbers will *not* match. Thus,
968 * ocfs2_cluster_lock() will not clear the PENDING set by
969 * ocfs2_prepare_downconvert().
970 */
971
972/* Unlocked version for ocfs2_locking_ast() */
973static void __lockres_clear_pending(struct ocfs2_lock_res *lockres,
974 unsigned int generation,
975 struct ocfs2_super *osb)
976{
977 assert_spin_locked(&lockres->l_lock);
978
979 /*
980 * The ast and locking functions can race us here. The winner
981 * will clear pending, the loser will not.
982 */
983 if (!(lockres->l_flags & OCFS2_LOCK_PENDING) ||
984 (lockres->l_pending_gen != generation))
985 return;
986
987 lockres_clear_flags(lockres, OCFS2_LOCK_PENDING);
988 lockres->l_pending_gen++;
989
990 /*
991 * The downconvert thread may have skipped us because we
992 * were PENDING. Wake it up.
993 */
994 if (lockres->l_flags & OCFS2_LOCK_BLOCKED)
995 ocfs2_wake_downconvert_thread(osb);
996}
997
998/* Locked version for callers of ocfs2_dlm_lock() */
999static void lockres_clear_pending(struct ocfs2_lock_res *lockres,
1000 unsigned int generation,
1001 struct ocfs2_super *osb)
1002{
1003 unsigned long flags;
1004
1005 spin_lock_irqsave(&lockres->l_lock, flags);
1006 __lockres_clear_pending(lockres, generation, osb);
1007 spin_unlock_irqrestore(&lockres->l_lock, flags);
1008}
1009
1010static unsigned int lockres_set_pending(struct ocfs2_lock_res *lockres)
1011{
1012 assert_spin_locked(&lockres->l_lock);
1013 BUG_ON(!(lockres->l_flags & OCFS2_LOCK_BUSY));
1014
1015 lockres_or_flags(lockres, OCFS2_LOCK_PENDING);
1016
1017 return lockres->l_pending_gen;
1018}
1019
1020static void ocfs2_blocking_ast(struct ocfs2_dlm_lksb *lksb, int level)
1021{
1022 struct ocfs2_lock_res *lockres = ocfs2_lksb_to_lock_res(lksb);
1023 struct ocfs2_super *osb = ocfs2_get_lockres_osb(lockres);
1024 int needs_downconvert;
1025 unsigned long flags;
1026
1027 BUG_ON(level <= DLM_LOCK_NL);
1028
1029 mlog(ML_BASTS, "BAST fired for lockres %s, blocking %d, level %d, "
1030 "type %s\n", lockres->l_name, level, lockres->l_level,
1031 ocfs2_lock_type_string(lockres->l_type));
1032
1033 /*
1034 * We can skip the bast for locks which don't enable caching -
1035 * they'll be dropped at the earliest possible time anyway.
1036 */
1037 if (lockres->l_flags & OCFS2_LOCK_NOCACHE)
1038 return;
1039
1040 spin_lock_irqsave(&lockres->l_lock, flags);
1041 needs_downconvert = ocfs2_generic_handle_bast(lockres, level);
1042 if (needs_downconvert)
1043 ocfs2_schedule_blocked_lock(osb, lockres);
1044 spin_unlock_irqrestore(&lockres->l_lock, flags);
1045
1046 wake_up(&lockres->l_event);
1047
1048 ocfs2_wake_downconvert_thread(osb);
1049}
1050
1051static void ocfs2_locking_ast(struct ocfs2_dlm_lksb *lksb)
1052{
1053 struct ocfs2_lock_res *lockres = ocfs2_lksb_to_lock_res(lksb);
1054 struct ocfs2_super *osb = ocfs2_get_lockres_osb(lockres);
1055 unsigned long flags;
1056 int status;
1057
1058 spin_lock_irqsave(&lockres->l_lock, flags);
1059
1060 status = ocfs2_dlm_lock_status(&lockres->l_lksb);
1061
1062 if (status == -EAGAIN) {
1063 lockres_clear_flags(lockres, OCFS2_LOCK_BUSY);
1064 goto out;
1065 }
1066
1067 if (status) {
1068 mlog(ML_ERROR, "lockres %s: lksb status value of %d!\n",
1069 lockres->l_name, status);
1070 spin_unlock_irqrestore(&lockres->l_lock, flags);
1071 return;
1072 }
1073
1074 mlog(ML_BASTS, "AST fired for lockres %s, action %d, unlock %d, "
1075 "level %d => %d\n", lockres->l_name, lockres->l_action,
1076 lockres->l_unlock_action, lockres->l_level, lockres->l_requested);
1077
1078 switch(lockres->l_action) {
1079 case OCFS2_AST_ATTACH:
1080 ocfs2_generic_handle_attach_action(lockres);
1081 lockres_clear_flags(lockres, OCFS2_LOCK_LOCAL);
1082 break;
1083 case OCFS2_AST_CONVERT:
1084 ocfs2_generic_handle_convert_action(lockres);
1085 break;
1086 case OCFS2_AST_DOWNCONVERT:
1087 ocfs2_generic_handle_downconvert_action(lockres);
1088 break;
1089 default:
1090 mlog(ML_ERROR, "lockres %s: AST fired with invalid action: %u, "
1091 "flags 0x%lx, unlock: %u\n",
1092 lockres->l_name, lockres->l_action, lockres->l_flags,
1093 lockres->l_unlock_action);
1094 BUG();
1095 }
1096out:
1097 /* set it to something invalid so if we get called again we
1098 * can catch it. */
1099 lockres->l_action = OCFS2_AST_INVALID;
1100
1101 /* Did we try to cancel this lock? Clear that state */
1102 if (lockres->l_unlock_action == OCFS2_UNLOCK_CANCEL_CONVERT)
1103 lockres->l_unlock_action = OCFS2_UNLOCK_INVALID;
1104
1105 /*
1106 * We may have beaten the locking functions here. We certainly
1107 * know that dlm_lock() has been called :-)
1108 * Because we can't have two lock calls in flight at once, we
1109 * can use lockres->l_pending_gen.
1110 */
1111 __lockres_clear_pending(lockres, lockres->l_pending_gen, osb);
1112
1113 wake_up(&lockres->l_event);
1114 spin_unlock_irqrestore(&lockres->l_lock, flags);
1115}
1116
1117static void ocfs2_unlock_ast(struct ocfs2_dlm_lksb *lksb, int error)
1118{
1119 struct ocfs2_lock_res *lockres = ocfs2_lksb_to_lock_res(lksb);
1120 unsigned long flags;
1121
1122 mlog(ML_BASTS, "UNLOCK AST fired for lockres %s, action = %d\n",
1123 lockres->l_name, lockres->l_unlock_action);
1124
1125 spin_lock_irqsave(&lockres->l_lock, flags);
1126 if (error) {
1127 mlog(ML_ERROR, "Dlm passes error %d for lock %s, "
1128 "unlock_action %d\n", error, lockres->l_name,
1129 lockres->l_unlock_action);
1130 spin_unlock_irqrestore(&lockres->l_lock, flags);
1131 return;
1132 }
1133
1134 switch(lockres->l_unlock_action) {
1135 case OCFS2_UNLOCK_CANCEL_CONVERT:
1136 mlog(0, "Cancel convert success for %s\n", lockres->l_name);
1137 lockres->l_action = OCFS2_AST_INVALID;
1138 /* Downconvert thread may have requeued this lock, we
1139 * need to wake it. */
1140 if (lockres->l_flags & OCFS2_LOCK_BLOCKED)
1141 ocfs2_wake_downconvert_thread(ocfs2_get_lockres_osb(lockres));
1142 break;
1143 case OCFS2_UNLOCK_DROP_LOCK:
1144 lockres->l_level = DLM_LOCK_IV;
1145 break;
1146 default:
1147 BUG();
1148 }
1149
1150 lockres_clear_flags(lockres, OCFS2_LOCK_BUSY);
1151 lockres->l_unlock_action = OCFS2_UNLOCK_INVALID;
1152 wake_up(&lockres->l_event);
1153 spin_unlock_irqrestore(&lockres->l_lock, flags);
1154}
1155
1156/*
1157 * This is the filesystem locking protocol. It provides the lock handling
1158 * hooks for the underlying DLM. It has a maximum version number.
1159 * The version number allows interoperability with systems running at
1160 * the same major number and an equal or smaller minor number.
1161 *
1162 * Whenever the filesystem does new things with locks (adds or removes a
1163 * lock, orders them differently, does different things underneath a lock),
1164 * the version must be changed. The protocol is negotiated when joining
1165 * the dlm domain. A node may join the domain if its major version is
1166 * identical to all other nodes and its minor version is greater than
1167 * or equal to all other nodes. When its minor version is greater than
1168 * the other nodes, it will run at the minor version specified by the
1169 * other nodes.
1170 *
1171 * If a locking change is made that will not be compatible with older
1172 * versions, the major number must be increased and the minor version set
1173 * to zero. If a change merely adds a behavior that can be disabled when
1174 * speaking to older versions, the minor version must be increased. If a
1175 * change adds a fully backwards compatible change (eg, LVB changes that
1176 * are just ignored by older versions), the version does not need to be
1177 * updated.
1178 */
1179static struct ocfs2_locking_protocol lproto = {
1180 .lp_max_version = {
1181 .pv_major = OCFS2_LOCKING_PROTOCOL_MAJOR,
1182 .pv_minor = OCFS2_LOCKING_PROTOCOL_MINOR,
1183 },
1184 .lp_lock_ast = ocfs2_locking_ast,
1185 .lp_blocking_ast = ocfs2_blocking_ast,
1186 .lp_unlock_ast = ocfs2_unlock_ast,
1187};
1188
1189void ocfs2_set_locking_protocol(void)
1190{
1191 ocfs2_stack_glue_set_max_proto_version(&lproto.lp_max_version);
1192}
1193
1194static inline void ocfs2_recover_from_dlm_error(struct ocfs2_lock_res *lockres,
1195 int convert)
1196{
1197 unsigned long flags;
1198
1199 spin_lock_irqsave(&lockres->l_lock, flags);
1200 lockres_clear_flags(lockres, OCFS2_LOCK_BUSY);
1201 lockres_clear_flags(lockres, OCFS2_LOCK_UPCONVERT_FINISHING);
1202 if (convert)
1203 lockres->l_action = OCFS2_AST_INVALID;
1204 else
1205 lockres->l_unlock_action = OCFS2_UNLOCK_INVALID;
1206 spin_unlock_irqrestore(&lockres->l_lock, flags);
1207
1208 wake_up(&lockres->l_event);
1209}
1210
1211/* Note: If we detect another process working on the lock (i.e.,
1212 * OCFS2_LOCK_BUSY), we'll bail out returning 0. It's up to the caller
1213 * to do the right thing in that case.
1214 */
1215static int ocfs2_lock_create(struct ocfs2_super *osb,
1216 struct ocfs2_lock_res *lockres,
1217 int level,
1218 u32 dlm_flags)
1219{
1220 int ret = 0;
1221 unsigned long flags;
1222 unsigned int gen;
1223
1224 mlog(0, "lock %s, level = %d, flags = %u\n", lockres->l_name, level,
1225 dlm_flags);
1226
1227 spin_lock_irqsave(&lockres->l_lock, flags);
1228 if ((lockres->l_flags & OCFS2_LOCK_ATTACHED) ||
1229 (lockres->l_flags & OCFS2_LOCK_BUSY)) {
1230 spin_unlock_irqrestore(&lockres->l_lock, flags);
1231 goto bail;
1232 }
1233
1234 lockres->l_action = OCFS2_AST_ATTACH;
1235 lockres->l_requested = level;
1236 lockres_or_flags(lockres, OCFS2_LOCK_BUSY);
1237 gen = lockres_set_pending(lockres);
1238 spin_unlock_irqrestore(&lockres->l_lock, flags);
1239
1240 ret = ocfs2_dlm_lock(osb->cconn,
1241 level,
1242 &lockres->l_lksb,
1243 dlm_flags,
1244 lockres->l_name,
1245 OCFS2_LOCK_ID_MAX_LEN - 1);
1246 lockres_clear_pending(lockres, gen, osb);
1247 if (ret) {
1248 ocfs2_log_dlm_error("ocfs2_dlm_lock", ret, lockres);
1249 ocfs2_recover_from_dlm_error(lockres, 1);
1250 }
1251
1252 mlog(0, "lock %s, return from ocfs2_dlm_lock\n", lockres->l_name);
1253
1254bail:
1255 return ret;
1256}
1257
1258static inline int ocfs2_check_wait_flag(struct ocfs2_lock_res *lockres,
1259 int flag)
1260{
1261 unsigned long flags;
1262 int ret;
1263
1264 spin_lock_irqsave(&lockres->l_lock, flags);
1265 ret = lockres->l_flags & flag;
1266 spin_unlock_irqrestore(&lockres->l_lock, flags);
1267
1268 return ret;
1269}
1270
1271static inline void ocfs2_wait_on_busy_lock(struct ocfs2_lock_res *lockres)
1272
1273{
1274 wait_event(lockres->l_event,
1275 !ocfs2_check_wait_flag(lockres, OCFS2_LOCK_BUSY));
1276}
1277
1278static inline void ocfs2_wait_on_refreshing_lock(struct ocfs2_lock_res *lockres)
1279
1280{
1281 wait_event(lockres->l_event,
1282 !ocfs2_check_wait_flag(lockres, OCFS2_LOCK_REFRESHING));
1283}
1284
1285/* predict what lock level we'll be dropping down to on behalf
1286 * of another node, and return true if the currently wanted
1287 * level will be compatible with it. */
1288static inline int ocfs2_may_continue_on_blocked_lock(struct ocfs2_lock_res *lockres,
1289 int wanted)
1290{
1291 BUG_ON(!(lockres->l_flags & OCFS2_LOCK_BLOCKED));
1292
1293 return wanted <= ocfs2_highest_compat_lock_level(lockres->l_blocking);
1294}
1295
1296static void ocfs2_init_mask_waiter(struct ocfs2_mask_waiter *mw)
1297{
1298 INIT_LIST_HEAD(&mw->mw_item);
1299 init_completion(&mw->mw_complete);
1300 ocfs2_init_start_time(mw);
1301}
1302
1303static int ocfs2_wait_for_mask(struct ocfs2_mask_waiter *mw)
1304{
1305 wait_for_completion(&mw->mw_complete);
1306 /* Re-arm the completion in case we want to wait on it again */
1307 INIT_COMPLETION(mw->mw_complete);
1308 return mw->mw_status;
1309}
1310
1311static void lockres_add_mask_waiter(struct ocfs2_lock_res *lockres,
1312 struct ocfs2_mask_waiter *mw,
1313 unsigned long mask,
1314 unsigned long goal)
1315{
1316 BUG_ON(!list_empty(&mw->mw_item));
1317
1318 assert_spin_locked(&lockres->l_lock);
1319
1320 list_add_tail(&mw->mw_item, &lockres->l_mask_waiters);
1321 mw->mw_mask = mask;
1322 mw->mw_goal = goal;
1323}
1324
1325/* returns 0 if the mw that was removed was already satisfied, -EBUSY
1326 * if the mask still hadn't reached its goal */
1327static int lockres_remove_mask_waiter(struct ocfs2_lock_res *lockres,
1328 struct ocfs2_mask_waiter *mw)
1329{
1330 unsigned long flags;
1331 int ret = 0;
1332
1333 spin_lock_irqsave(&lockres->l_lock, flags);
1334 if (!list_empty(&mw->mw_item)) {
1335 if ((lockres->l_flags & mw->mw_mask) != mw->mw_goal)
1336 ret = -EBUSY;
1337
1338 list_del_init(&mw->mw_item);
1339 init_completion(&mw->mw_complete);
1340 }
1341 spin_unlock_irqrestore(&lockres->l_lock, flags);
1342
1343 return ret;
1344
1345}
1346
1347static int ocfs2_wait_for_mask_interruptible(struct ocfs2_mask_waiter *mw,
1348 struct ocfs2_lock_res *lockres)
1349{
1350 int ret;
1351
1352 ret = wait_for_completion_interruptible(&mw->mw_complete);
1353 if (ret)
1354 lockres_remove_mask_waiter(lockres, mw);
1355 else
1356 ret = mw->mw_status;
1357 /* Re-arm the completion in case we want to wait on it again */
1358 INIT_COMPLETION(mw->mw_complete);
1359 return ret;
1360}
1361
1362static int __ocfs2_cluster_lock(struct ocfs2_super *osb,
1363 struct ocfs2_lock_res *lockres,
1364 int level,
1365 u32 lkm_flags,
1366 int arg_flags,
1367 int l_subclass,
1368 unsigned long caller_ip)
1369{
1370 struct ocfs2_mask_waiter mw;
1371 int wait, catch_signals = !(osb->s_mount_opt & OCFS2_MOUNT_NOINTR);
1372 int ret = 0; /* gcc doesn't realize wait = 1 guarantees ret is set */
1373 unsigned long flags;
1374 unsigned int gen;
1375 int noqueue_attempted = 0;
1376
1377 ocfs2_init_mask_waiter(&mw);
1378
1379 if (lockres->l_ops->flags & LOCK_TYPE_USES_LVB)
1380 lkm_flags |= DLM_LKF_VALBLK;
1381
1382again:
1383 wait = 0;
1384
1385 spin_lock_irqsave(&lockres->l_lock, flags);
1386
1387 if (catch_signals && signal_pending(current)) {
1388 ret = -ERESTARTSYS;
1389 goto unlock;
1390 }
1391
1392 mlog_bug_on_msg(lockres->l_flags & OCFS2_LOCK_FREEING,
1393 "Cluster lock called on freeing lockres %s! flags "
1394 "0x%lx\n", lockres->l_name, lockres->l_flags);
1395
1396 /* We only compare against the currently granted level
1397 * here. If the lock is blocked waiting on a downconvert,
1398 * we'll get caught below. */
1399 if (lockres->l_flags & OCFS2_LOCK_BUSY &&
1400 level > lockres->l_level) {
1401 /* is someone sitting in dlm_lock? If so, wait on
1402 * them. */
1403 lockres_add_mask_waiter(lockres, &mw, OCFS2_LOCK_BUSY, 0);
1404 wait = 1;
1405 goto unlock;
1406 }
1407
1408 if (lockres->l_flags & OCFS2_LOCK_UPCONVERT_FINISHING) {
1409 /*
1410 * We've upconverted. If the lock now has a level we can
1411 * work with, we take it. If, however, the lock is not at the
1412 * required level, we go thru the full cycle. One way this could
1413 * happen is if a process requesting an upconvert to PR is
1414 * closely followed by another requesting upconvert to an EX.
1415 * If the process requesting EX lands here, we want it to
1416 * continue attempting to upconvert and let the process
1417 * requesting PR take the lock.
1418 * If multiple processes request upconvert to PR, the first one
1419 * here will take the lock. The others will have to go thru the
1420 * OCFS2_LOCK_BLOCKED check to ensure that there is no pending
1421 * downconvert request.
1422 */
1423 if (level <= lockres->l_level)
1424 goto update_holders;
1425 }
1426
1427 if (lockres->l_flags & OCFS2_LOCK_BLOCKED &&
1428 !ocfs2_may_continue_on_blocked_lock(lockres, level)) {
1429 /* is the lock is currently blocked on behalf of
1430 * another node */
1431 lockres_add_mask_waiter(lockres, &mw, OCFS2_LOCK_BLOCKED, 0);
1432 wait = 1;
1433 goto unlock;
1434 }
1435
1436 if (level > lockres->l_level) {
1437 if (noqueue_attempted > 0) {
1438 ret = -EAGAIN;
1439 goto unlock;
1440 }
1441 if (lkm_flags & DLM_LKF_NOQUEUE)
1442 noqueue_attempted = 1;
1443
1444 if (lockres->l_action != OCFS2_AST_INVALID)
1445 mlog(ML_ERROR, "lockres %s has action %u pending\n",
1446 lockres->l_name, lockres->l_action);
1447
1448 if (!(lockres->l_flags & OCFS2_LOCK_ATTACHED)) {
1449 lockres->l_action = OCFS2_AST_ATTACH;
1450 lkm_flags &= ~DLM_LKF_CONVERT;
1451 } else {
1452 lockres->l_action = OCFS2_AST_CONVERT;
1453 lkm_flags |= DLM_LKF_CONVERT;
1454 }
1455
1456 lockres->l_requested = level;
1457 lockres_or_flags(lockres, OCFS2_LOCK_BUSY);
1458 gen = lockres_set_pending(lockres);
1459 spin_unlock_irqrestore(&lockres->l_lock, flags);
1460
1461 BUG_ON(level == DLM_LOCK_IV);
1462 BUG_ON(level == DLM_LOCK_NL);
1463
1464 mlog(ML_BASTS, "lockres %s, convert from %d to %d\n",
1465 lockres->l_name, lockres->l_level, level);
1466
1467 /* call dlm_lock to upgrade lock now */
1468 ret = ocfs2_dlm_lock(osb->cconn,
1469 level,
1470 &lockres->l_lksb,
1471 lkm_flags,
1472 lockres->l_name,
1473 OCFS2_LOCK_ID_MAX_LEN - 1);
1474 lockres_clear_pending(lockres, gen, osb);
1475 if (ret) {
1476 if (!(lkm_flags & DLM_LKF_NOQUEUE) ||
1477 (ret != -EAGAIN)) {
1478 ocfs2_log_dlm_error("ocfs2_dlm_lock",
1479 ret, lockres);
1480 }
1481 ocfs2_recover_from_dlm_error(lockres, 1);
1482 goto out;
1483 }
1484
1485 mlog(0, "lock %s, successful return from ocfs2_dlm_lock\n",
1486 lockres->l_name);
1487
1488 /* At this point we've gone inside the dlm and need to
1489 * complete our work regardless. */
1490 catch_signals = 0;
1491
1492 /* wait for busy to clear and carry on */
1493 goto again;
1494 }
1495
1496update_holders:
1497 /* Ok, if we get here then we're good to go. */
1498 ocfs2_inc_holders(lockres, level);
1499
1500 ret = 0;
1501unlock:
1502 lockres_clear_flags(lockres, OCFS2_LOCK_UPCONVERT_FINISHING);
1503
1504 spin_unlock_irqrestore(&lockres->l_lock, flags);
1505out:
1506 /*
1507 * This is helping work around a lock inversion between the page lock
1508 * and dlm locks. One path holds the page lock while calling aops
1509 * which block acquiring dlm locks. The voting thread holds dlm
1510 * locks while acquiring page locks while down converting data locks.
1511 * This block is helping an aop path notice the inversion and back
1512 * off to unlock its page lock before trying the dlm lock again.
1513 */
1514 if (wait && arg_flags & OCFS2_LOCK_NONBLOCK &&
1515 mw.mw_mask & (OCFS2_LOCK_BUSY|OCFS2_LOCK_BLOCKED)) {
1516 wait = 0;
1517 if (lockres_remove_mask_waiter(lockres, &mw))
1518 ret = -EAGAIN;
1519 else
1520 goto again;
1521 }
1522 if (wait) {
1523 ret = ocfs2_wait_for_mask(&mw);
1524 if (ret == 0)
1525 goto again;
1526 mlog_errno(ret);
1527 }
1528 ocfs2_update_lock_stats(lockres, level, &mw, ret);
1529
1530#ifdef CONFIG_DEBUG_LOCK_ALLOC
1531 if (!ret && lockres->l_lockdep_map.key != NULL) {
1532 if (level == DLM_LOCK_PR)
1533 rwsem_acquire_read(&lockres->l_lockdep_map, l_subclass,
1534 !!(arg_flags & OCFS2_META_LOCK_NOQUEUE),
1535 caller_ip);
1536 else
1537 rwsem_acquire(&lockres->l_lockdep_map, l_subclass,
1538 !!(arg_flags & OCFS2_META_LOCK_NOQUEUE),
1539 caller_ip);
1540 }
1541#endif
1542 return ret;
1543}
1544
1545static inline int ocfs2_cluster_lock(struct ocfs2_super *osb,
1546 struct ocfs2_lock_res *lockres,
1547 int level,
1548 u32 lkm_flags,
1549 int arg_flags)
1550{
1551 return __ocfs2_cluster_lock(osb, lockres, level, lkm_flags, arg_flags,
1552 0, _RET_IP_);
1553}
1554
1555
1556static void __ocfs2_cluster_unlock(struct ocfs2_super *osb,
1557 struct ocfs2_lock_res *lockres,
1558 int level,
1559 unsigned long caller_ip)
1560{
1561 unsigned long flags;
1562
1563 spin_lock_irqsave(&lockres->l_lock, flags);
1564 ocfs2_dec_holders(lockres, level);
1565 ocfs2_downconvert_on_unlock(osb, lockres);
1566 spin_unlock_irqrestore(&lockres->l_lock, flags);
1567#ifdef CONFIG_DEBUG_LOCK_ALLOC
1568 if (lockres->l_lockdep_map.key != NULL)
1569 rwsem_release(&lockres->l_lockdep_map, 1, caller_ip);
1570#endif
1571}
1572
1573static int ocfs2_create_new_lock(struct ocfs2_super *osb,
1574 struct ocfs2_lock_res *lockres,
1575 int ex,
1576 int local)
1577{
1578 int level = ex ? DLM_LOCK_EX : DLM_LOCK_PR;
1579 unsigned long flags;
1580 u32 lkm_flags = local ? DLM_LKF_LOCAL : 0;
1581
1582 spin_lock_irqsave(&lockres->l_lock, flags);
1583 BUG_ON(lockres->l_flags & OCFS2_LOCK_ATTACHED);
1584 lockres_or_flags(lockres, OCFS2_LOCK_LOCAL);
1585 spin_unlock_irqrestore(&lockres->l_lock, flags);
1586
1587 return ocfs2_lock_create(osb, lockres, level, lkm_flags);
1588}
1589
1590/* Grants us an EX lock on the data and metadata resources, skipping
1591 * the normal cluster directory lookup. Use this ONLY on newly created
1592 * inodes which other nodes can't possibly see, and which haven't been
1593 * hashed in the inode hash yet. This can give us a good performance
1594 * increase as it'll skip the network broadcast normally associated
1595 * with creating a new lock resource. */
1596int ocfs2_create_new_inode_locks(struct inode *inode)
1597{
1598 int ret;
1599 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1600
1601 BUG_ON(!inode);
1602 BUG_ON(!ocfs2_inode_is_new(inode));
1603
1604 mlog(0, "Inode %llu\n", (unsigned long long)OCFS2_I(inode)->ip_blkno);
1605
1606 /* NOTE: That we don't increment any of the holder counts, nor
1607 * do we add anything to a journal handle. Since this is
1608 * supposed to be a new inode which the cluster doesn't know
1609 * about yet, there is no need to. As far as the LVB handling
1610 * is concerned, this is basically like acquiring an EX lock
1611 * on a resource which has an invalid one -- we'll set it
1612 * valid when we release the EX. */
1613
1614 ret = ocfs2_create_new_lock(osb, &OCFS2_I(inode)->ip_rw_lockres, 1, 1);
1615 if (ret) {
1616 mlog_errno(ret);
1617 goto bail;
1618 }
1619
1620 /*
1621 * We don't want to use DLM_LKF_LOCAL on a meta data lock as they
1622 * don't use a generation in their lock names.
1623 */
1624 ret = ocfs2_create_new_lock(osb, &OCFS2_I(inode)->ip_inode_lockres, 1, 0);
1625 if (ret) {
1626 mlog_errno(ret);
1627 goto bail;
1628 }
1629
1630 ret = ocfs2_create_new_lock(osb, &OCFS2_I(inode)->ip_open_lockres, 0, 0);
1631 if (ret) {
1632 mlog_errno(ret);
1633 goto bail;
1634 }
1635
1636bail:
1637 return ret;
1638}
1639
1640int ocfs2_rw_lock(struct inode *inode, int write)
1641{
1642 int status, level;
1643 struct ocfs2_lock_res *lockres;
1644 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1645
1646 BUG_ON(!inode);
1647
1648 mlog(0, "inode %llu take %s RW lock\n",
1649 (unsigned long long)OCFS2_I(inode)->ip_blkno,
1650 write ? "EXMODE" : "PRMODE");
1651
1652 if (ocfs2_mount_local(osb))
1653 return 0;
1654
1655 lockres = &OCFS2_I(inode)->ip_rw_lockres;
1656
1657 level = write ? DLM_LOCK_EX : DLM_LOCK_PR;
1658
1659 status = ocfs2_cluster_lock(OCFS2_SB(inode->i_sb), lockres, level, 0,
1660 0);
1661 if (status < 0)
1662 mlog_errno(status);
1663
1664 return status;
1665}
1666
1667void ocfs2_rw_unlock(struct inode *inode, int write)
1668{
1669 int level = write ? DLM_LOCK_EX : DLM_LOCK_PR;
1670 struct ocfs2_lock_res *lockres = &OCFS2_I(inode)->ip_rw_lockres;
1671 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1672
1673 mlog(0, "inode %llu drop %s RW lock\n",
1674 (unsigned long long)OCFS2_I(inode)->ip_blkno,
1675 write ? "EXMODE" : "PRMODE");
1676
1677 if (!ocfs2_mount_local(osb))
1678 ocfs2_cluster_unlock(OCFS2_SB(inode->i_sb), lockres, level);
1679}
1680
1681/*
1682 * ocfs2_open_lock always get PR mode lock.
1683 */
1684int ocfs2_open_lock(struct inode *inode)
1685{
1686 int status = 0;
1687 struct ocfs2_lock_res *lockres;
1688 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1689
1690 BUG_ON(!inode);
1691
1692 mlog(0, "inode %llu take PRMODE open lock\n",
1693 (unsigned long long)OCFS2_I(inode)->ip_blkno);
1694
1695 if (ocfs2_mount_local(osb))
1696 goto out;
1697
1698 lockres = &OCFS2_I(inode)->ip_open_lockres;
1699
1700 status = ocfs2_cluster_lock(OCFS2_SB(inode->i_sb), lockres,
1701 DLM_LOCK_PR, 0, 0);
1702 if (status < 0)
1703 mlog_errno(status);
1704
1705out:
1706 return status;
1707}
1708
1709int ocfs2_try_open_lock(struct inode *inode, int write)
1710{
1711 int status = 0, level;
1712 struct ocfs2_lock_res *lockres;
1713 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1714
1715 BUG_ON(!inode);
1716
1717 mlog(0, "inode %llu try to take %s open lock\n",
1718 (unsigned long long)OCFS2_I(inode)->ip_blkno,
1719 write ? "EXMODE" : "PRMODE");
1720
1721 if (ocfs2_mount_local(osb))
1722 goto out;
1723
1724 lockres = &OCFS2_I(inode)->ip_open_lockres;
1725
1726 level = write ? DLM_LOCK_EX : DLM_LOCK_PR;
1727
1728 /*
1729 * The file system may already holding a PRMODE/EXMODE open lock.
1730 * Since we pass DLM_LKF_NOQUEUE, the request won't block waiting on
1731 * other nodes and the -EAGAIN will indicate to the caller that
1732 * this inode is still in use.
1733 */
1734 status = ocfs2_cluster_lock(OCFS2_SB(inode->i_sb), lockres,
1735 level, DLM_LKF_NOQUEUE, 0);
1736
1737out:
1738 return status;
1739}
1740
1741/*
1742 * ocfs2_open_unlock unlock PR and EX mode open locks.
1743 */
1744void ocfs2_open_unlock(struct inode *inode)
1745{
1746 struct ocfs2_lock_res *lockres = &OCFS2_I(inode)->ip_open_lockres;
1747 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1748
1749 mlog(0, "inode %llu drop open lock\n",
1750 (unsigned long long)OCFS2_I(inode)->ip_blkno);
1751
1752 if (ocfs2_mount_local(osb))
1753 goto out;
1754
1755 if(lockres->l_ro_holders)
1756 ocfs2_cluster_unlock(OCFS2_SB(inode->i_sb), lockres,
1757 DLM_LOCK_PR);
1758 if(lockres->l_ex_holders)
1759 ocfs2_cluster_unlock(OCFS2_SB(inode->i_sb), lockres,
1760 DLM_LOCK_EX);
1761
1762out:
1763 return;
1764}
1765
1766static int ocfs2_flock_handle_signal(struct ocfs2_lock_res *lockres,
1767 int level)
1768{
1769 int ret;
1770 struct ocfs2_super *osb = ocfs2_get_lockres_osb(lockres);
1771 unsigned long flags;
1772 struct ocfs2_mask_waiter mw;
1773
1774 ocfs2_init_mask_waiter(&mw);
1775
1776retry_cancel:
1777 spin_lock_irqsave(&lockres->l_lock, flags);
1778 if (lockres->l_flags & OCFS2_LOCK_BUSY) {
1779 ret = ocfs2_prepare_cancel_convert(osb, lockres);
1780 if (ret) {
1781 spin_unlock_irqrestore(&lockres->l_lock, flags);
1782 ret = ocfs2_cancel_convert(osb, lockres);
1783 if (ret < 0) {
1784 mlog_errno(ret);
1785 goto out;
1786 }
1787 goto retry_cancel;
1788 }
1789 lockres_add_mask_waiter(lockres, &mw, OCFS2_LOCK_BUSY, 0);
1790 spin_unlock_irqrestore(&lockres->l_lock, flags);
1791
1792 ocfs2_wait_for_mask(&mw);
1793 goto retry_cancel;
1794 }
1795
1796 ret = -ERESTARTSYS;
1797 /*
1798 * We may still have gotten the lock, in which case there's no
1799 * point to restarting the syscall.
1800 */
1801 if (lockres->l_level == level)
1802 ret = 0;
1803
1804 mlog(0, "Cancel returning %d. flags: 0x%lx, level: %d, act: %d\n", ret,
1805 lockres->l_flags, lockres->l_level, lockres->l_action);
1806
1807 spin_unlock_irqrestore(&lockres->l_lock, flags);
1808
1809out:
1810 return ret;
1811}
1812
1813/*
1814 * ocfs2_file_lock() and ocfs2_file_unlock() map to a single pair of
1815 * flock() calls. The locking approach this requires is sufficiently
1816 * different from all other cluster lock types that we implement a
1817 * separate path to the "low-level" dlm calls. In particular:
1818 *
1819 * - No optimization of lock levels is done - we take at exactly
1820 * what's been requested.
1821 *
1822 * - No lock caching is employed. We immediately downconvert to
1823 * no-lock at unlock time. This also means flock locks never go on
1824 * the blocking list).
1825 *
1826 * - Since userspace can trivially deadlock itself with flock, we make
1827 * sure to allow cancellation of a misbehaving applications flock()
1828 * request.
1829 *
1830 * - Access to any flock lockres doesn't require concurrency, so we
1831 * can simplify the code by requiring the caller to guarantee
1832 * serialization of dlmglue flock calls.
1833 */
1834int ocfs2_file_lock(struct file *file, int ex, int trylock)
1835{
1836 int ret, level = ex ? DLM_LOCK_EX : DLM_LOCK_PR;
1837 unsigned int lkm_flags = trylock ? DLM_LKF_NOQUEUE : 0;
1838 unsigned long flags;
1839 struct ocfs2_file_private *fp = file->private_data;
1840 struct ocfs2_lock_res *lockres = &fp->fp_flock;
1841 struct ocfs2_super *osb = OCFS2_SB(file->f_mapping->host->i_sb);
1842 struct ocfs2_mask_waiter mw;
1843
1844 ocfs2_init_mask_waiter(&mw);
1845
1846 if ((lockres->l_flags & OCFS2_LOCK_BUSY) ||
1847 (lockres->l_level > DLM_LOCK_NL)) {
1848 mlog(ML_ERROR,
1849 "File lock \"%s\" has busy or locked state: flags: 0x%lx, "
1850 "level: %u\n", lockres->l_name, lockres->l_flags,
1851 lockres->l_level);
1852 return -EINVAL;
1853 }
1854
1855 spin_lock_irqsave(&lockres->l_lock, flags);
1856 if (!(lockres->l_flags & OCFS2_LOCK_ATTACHED)) {
1857 lockres_add_mask_waiter(lockres, &mw, OCFS2_LOCK_BUSY, 0);
1858 spin_unlock_irqrestore(&lockres->l_lock, flags);
1859
1860 /*
1861 * Get the lock at NLMODE to start - that way we
1862 * can cancel the upconvert request if need be.
1863 */
1864 ret = ocfs2_lock_create(osb, lockres, DLM_LOCK_NL, 0);
1865 if (ret < 0) {
1866 mlog_errno(ret);
1867 goto out;
1868 }
1869
1870 ret = ocfs2_wait_for_mask(&mw);
1871 if (ret) {
1872 mlog_errno(ret);
1873 goto out;
1874 }
1875 spin_lock_irqsave(&lockres->l_lock, flags);
1876 }
1877
1878 lockres->l_action = OCFS2_AST_CONVERT;
1879 lkm_flags |= DLM_LKF_CONVERT;
1880 lockres->l_requested = level;
1881 lockres_or_flags(lockres, OCFS2_LOCK_BUSY);
1882
1883 lockres_add_mask_waiter(lockres, &mw, OCFS2_LOCK_BUSY, 0);
1884 spin_unlock_irqrestore(&lockres->l_lock, flags);
1885
1886 ret = ocfs2_dlm_lock(osb->cconn, level, &lockres->l_lksb, lkm_flags,
1887 lockres->l_name, OCFS2_LOCK_ID_MAX_LEN - 1);
1888 if (ret) {
1889 if (!trylock || (ret != -EAGAIN)) {
1890 ocfs2_log_dlm_error("ocfs2_dlm_lock", ret, lockres);
1891 ret = -EINVAL;
1892 }
1893
1894 ocfs2_recover_from_dlm_error(lockres, 1);
1895 lockres_remove_mask_waiter(lockres, &mw);
1896 goto out;
1897 }
1898
1899 ret = ocfs2_wait_for_mask_interruptible(&mw, lockres);
1900 if (ret == -ERESTARTSYS) {
1901 /*
1902 * Userspace can cause deadlock itself with
1903 * flock(). Current behavior locally is to allow the
1904 * deadlock, but abort the system call if a signal is
1905 * received. We follow this example, otherwise a
1906 * poorly written program could sit in kernel until
1907 * reboot.
1908 *
1909 * Handling this is a bit more complicated for Ocfs2
1910 * though. We can't exit this function with an
1911 * outstanding lock request, so a cancel convert is
1912 * required. We intentionally overwrite 'ret' - if the
1913 * cancel fails and the lock was granted, it's easier
1914 * to just bubble success back up to the user.
1915 */
1916 ret = ocfs2_flock_handle_signal(lockres, level);
1917 } else if (!ret && (level > lockres->l_level)) {
1918 /* Trylock failed asynchronously */
1919 BUG_ON(!trylock);
1920 ret = -EAGAIN;
1921 }
1922
1923out:
1924
1925 mlog(0, "Lock: \"%s\" ex: %d, trylock: %d, returns: %d\n",
1926 lockres->l_name, ex, trylock, ret);
1927 return ret;
1928}
1929
1930void ocfs2_file_unlock(struct file *file)
1931{
1932 int ret;
1933 unsigned int gen;
1934 unsigned long flags;
1935 struct ocfs2_file_private *fp = file->private_data;
1936 struct ocfs2_lock_res *lockres = &fp->fp_flock;
1937 struct ocfs2_super *osb = OCFS2_SB(file->f_mapping->host->i_sb);
1938 struct ocfs2_mask_waiter mw;
1939
1940 ocfs2_init_mask_waiter(&mw);
1941
1942 if (!(lockres->l_flags & OCFS2_LOCK_ATTACHED))
1943 return;
1944
1945 if (lockres->l_level == DLM_LOCK_NL)
1946 return;
1947
1948 mlog(0, "Unlock: \"%s\" flags: 0x%lx, level: %d, act: %d\n",
1949 lockres->l_name, lockres->l_flags, lockres->l_level,
1950 lockres->l_action);
1951
1952 spin_lock_irqsave(&lockres->l_lock, flags);
1953 /*
1954 * Fake a blocking ast for the downconvert code.
1955 */
1956 lockres_or_flags(lockres, OCFS2_LOCK_BLOCKED);
1957 lockres->l_blocking = DLM_LOCK_EX;
1958
1959 gen = ocfs2_prepare_downconvert(lockres, DLM_LOCK_NL);
1960 lockres_add_mask_waiter(lockres, &mw, OCFS2_LOCK_BUSY, 0);
1961 spin_unlock_irqrestore(&lockres->l_lock, flags);
1962
1963 ret = ocfs2_downconvert_lock(osb, lockres, DLM_LOCK_NL, 0, gen);
1964 if (ret) {
1965 mlog_errno(ret);
1966 return;
1967 }
1968
1969 ret = ocfs2_wait_for_mask(&mw);
1970 if (ret)
1971 mlog_errno(ret);
1972}
1973
1974static void ocfs2_downconvert_on_unlock(struct ocfs2_super *osb,
1975 struct ocfs2_lock_res *lockres)
1976{
1977 int kick = 0;
1978
1979 /* If we know that another node is waiting on our lock, kick
1980 * the downconvert thread * pre-emptively when we reach a release
1981 * condition. */
1982 if (lockres->l_flags & OCFS2_LOCK_BLOCKED) {
1983 switch(lockres->l_blocking) {
1984 case DLM_LOCK_EX:
1985 if (!lockres->l_ex_holders && !lockres->l_ro_holders)
1986 kick = 1;
1987 break;
1988 case DLM_LOCK_PR:
1989 if (!lockres->l_ex_holders)
1990 kick = 1;
1991 break;
1992 default:
1993 BUG();
1994 }
1995 }
1996
1997 if (kick)
1998 ocfs2_wake_downconvert_thread(osb);
1999}
2000
2001#define OCFS2_SEC_BITS 34
2002#define OCFS2_SEC_SHIFT (64 - 34)
2003#define OCFS2_NSEC_MASK ((1ULL << OCFS2_SEC_SHIFT) - 1)
2004
2005/* LVB only has room for 64 bits of time here so we pack it for
2006 * now. */
2007static u64 ocfs2_pack_timespec(struct timespec *spec)
2008{
2009 u64 res;
2010 u64 sec = spec->tv_sec;
2011 u32 nsec = spec->tv_nsec;
2012
2013 res = (sec << OCFS2_SEC_SHIFT) | (nsec & OCFS2_NSEC_MASK);
2014
2015 return res;
2016}
2017
2018/* Call this with the lockres locked. I am reasonably sure we don't
2019 * need ip_lock in this function as anyone who would be changing those
2020 * values is supposed to be blocked in ocfs2_inode_lock right now. */
2021static void __ocfs2_stuff_meta_lvb(struct inode *inode)
2022{
2023 struct ocfs2_inode_info *oi = OCFS2_I(inode);
2024 struct ocfs2_lock_res *lockres = &oi->ip_inode_lockres;
2025 struct ocfs2_meta_lvb *lvb;
2026
2027 lvb = ocfs2_dlm_lvb(&lockres->l_lksb);
2028
2029 /*
2030 * Invalidate the LVB of a deleted inode - this way other
2031 * nodes are forced to go to disk and discover the new inode
2032 * status.
2033 */
2034 if (oi->ip_flags & OCFS2_INODE_DELETED) {
2035 lvb->lvb_version = 0;
2036 goto out;
2037 }
2038
2039 lvb->lvb_version = OCFS2_LVB_VERSION;
2040 lvb->lvb_isize = cpu_to_be64(i_size_read(inode));
2041 lvb->lvb_iclusters = cpu_to_be32(oi->ip_clusters);
2042 lvb->lvb_iuid = cpu_to_be32(inode->i_uid);
2043 lvb->lvb_igid = cpu_to_be32(inode->i_gid);
2044 lvb->lvb_imode = cpu_to_be16(inode->i_mode);
2045 lvb->lvb_inlink = cpu_to_be16(inode->i_nlink);
2046 lvb->lvb_iatime_packed =
2047 cpu_to_be64(ocfs2_pack_timespec(&inode->i_atime));
2048 lvb->lvb_ictime_packed =
2049 cpu_to_be64(ocfs2_pack_timespec(&inode->i_ctime));
2050 lvb->lvb_imtime_packed =
2051 cpu_to_be64(ocfs2_pack_timespec(&inode->i_mtime));
2052 lvb->lvb_iattr = cpu_to_be32(oi->ip_attr);
2053 lvb->lvb_idynfeatures = cpu_to_be16(oi->ip_dyn_features);
2054 lvb->lvb_igeneration = cpu_to_be32(inode->i_generation);
2055
2056out:
2057 mlog_meta_lvb(0, lockres);
2058}
2059
2060static void ocfs2_unpack_timespec(struct timespec *spec,
2061 u64 packed_time)
2062{
2063 spec->tv_sec = packed_time >> OCFS2_SEC_SHIFT;
2064 spec->tv_nsec = packed_time & OCFS2_NSEC_MASK;
2065}
2066
2067static void ocfs2_refresh_inode_from_lvb(struct inode *inode)
2068{
2069 struct ocfs2_inode_info *oi = OCFS2_I(inode);
2070 struct ocfs2_lock_res *lockres = &oi->ip_inode_lockres;
2071 struct ocfs2_meta_lvb *lvb;
2072
2073 mlog_meta_lvb(0, lockres);
2074
2075 lvb = ocfs2_dlm_lvb(&lockres->l_lksb);
2076
2077 /* We're safe here without the lockres lock... */
2078 spin_lock(&oi->ip_lock);
2079 oi->ip_clusters = be32_to_cpu(lvb->lvb_iclusters);
2080 i_size_write(inode, be64_to_cpu(lvb->lvb_isize));
2081
2082 oi->ip_attr = be32_to_cpu(lvb->lvb_iattr);
2083 oi->ip_dyn_features = be16_to_cpu(lvb->lvb_idynfeatures);
2084 ocfs2_set_inode_flags(inode);
2085
2086 /* fast-symlinks are a special case */
2087 if (S_ISLNK(inode->i_mode) && !oi->ip_clusters)
2088 inode->i_blocks = 0;
2089 else
2090 inode->i_blocks = ocfs2_inode_sector_count(inode);
2091
2092 inode->i_uid = be32_to_cpu(lvb->lvb_iuid);
2093 inode->i_gid = be32_to_cpu(lvb->lvb_igid);
2094 inode->i_mode = be16_to_cpu(lvb->lvb_imode);
2095 inode->i_nlink = be16_to_cpu(lvb->lvb_inlink);
2096 ocfs2_unpack_timespec(&inode->i_atime,
2097 be64_to_cpu(lvb->lvb_iatime_packed));
2098 ocfs2_unpack_timespec(&inode->i_mtime,
2099 be64_to_cpu(lvb->lvb_imtime_packed));
2100 ocfs2_unpack_timespec(&inode->i_ctime,
2101 be64_to_cpu(lvb->lvb_ictime_packed));
2102 spin_unlock(&oi->ip_lock);
2103}
2104
2105static inline int ocfs2_meta_lvb_is_trustable(struct inode *inode,
2106 struct ocfs2_lock_res *lockres)
2107{
2108 struct ocfs2_meta_lvb *lvb = ocfs2_dlm_lvb(&lockres->l_lksb);
2109
2110 if (ocfs2_dlm_lvb_valid(&lockres->l_lksb)
2111 && lvb->lvb_version == OCFS2_LVB_VERSION
2112 && be32_to_cpu(lvb->lvb_igeneration) == inode->i_generation)
2113 return 1;
2114 return 0;
2115}
2116
2117/* Determine whether a lock resource needs to be refreshed, and
2118 * arbitrate who gets to refresh it.
2119 *
2120 * 0 means no refresh needed.
2121 *
2122 * > 0 means you need to refresh this and you MUST call
2123 * ocfs2_complete_lock_res_refresh afterwards. */
2124static int ocfs2_should_refresh_lock_res(struct ocfs2_lock_res *lockres)
2125{
2126 unsigned long flags;
2127 int status = 0;
2128
2129refresh_check:
2130 spin_lock_irqsave(&lockres->l_lock, flags);
2131 if (!(lockres->l_flags & OCFS2_LOCK_NEEDS_REFRESH)) {
2132 spin_unlock_irqrestore(&lockres->l_lock, flags);
2133 goto bail;
2134 }
2135
2136 if (lockres->l_flags & OCFS2_LOCK_REFRESHING) {
2137 spin_unlock_irqrestore(&lockres->l_lock, flags);
2138
2139 ocfs2_wait_on_refreshing_lock(lockres);
2140 goto refresh_check;
2141 }
2142
2143 /* Ok, I'll be the one to refresh this lock. */
2144 lockres_or_flags(lockres, OCFS2_LOCK_REFRESHING);
2145 spin_unlock_irqrestore(&lockres->l_lock, flags);
2146
2147 status = 1;
2148bail:
2149 mlog(0, "status %d\n", status);
2150 return status;
2151}
2152
2153/* If status is non zero, I'll mark it as not being in refresh
2154 * anymroe, but i won't clear the needs refresh flag. */
2155static inline void ocfs2_complete_lock_res_refresh(struct ocfs2_lock_res *lockres,
2156 int status)
2157{
2158 unsigned long flags;
2159
2160 spin_lock_irqsave(&lockres->l_lock, flags);
2161 lockres_clear_flags(lockres, OCFS2_LOCK_REFRESHING);
2162 if (!status)
2163 lockres_clear_flags(lockres, OCFS2_LOCK_NEEDS_REFRESH);
2164 spin_unlock_irqrestore(&lockres->l_lock, flags);
2165
2166 wake_up(&lockres->l_event);
2167}
2168
2169/* may or may not return a bh if it went to disk. */
2170static int ocfs2_inode_lock_update(struct inode *inode,
2171 struct buffer_head **bh)
2172{
2173 int status = 0;
2174 struct ocfs2_inode_info *oi = OCFS2_I(inode);
2175 struct ocfs2_lock_res *lockres = &oi->ip_inode_lockres;
2176 struct ocfs2_dinode *fe;
2177 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
2178
2179 if (ocfs2_mount_local(osb))
2180 goto bail;
2181
2182 spin_lock(&oi->ip_lock);
2183 if (oi->ip_flags & OCFS2_INODE_DELETED) {
2184 mlog(0, "Orphaned inode %llu was deleted while we "
2185 "were waiting on a lock. ip_flags = 0x%x\n",
2186 (unsigned long long)oi->ip_blkno, oi->ip_flags);
2187 spin_unlock(&oi->ip_lock);
2188 status = -ENOENT;
2189 goto bail;
2190 }
2191 spin_unlock(&oi->ip_lock);
2192
2193 if (!ocfs2_should_refresh_lock_res(lockres))
2194 goto bail;
2195
2196 /* This will discard any caching information we might have had
2197 * for the inode metadata. */
2198 ocfs2_metadata_cache_purge(INODE_CACHE(inode));
2199
2200 ocfs2_extent_map_trunc(inode, 0);
2201
2202 if (ocfs2_meta_lvb_is_trustable(inode, lockres)) {
2203 mlog(0, "Trusting LVB on inode %llu\n",
2204 (unsigned long long)oi->ip_blkno);
2205 ocfs2_refresh_inode_from_lvb(inode);
2206 } else {
2207 /* Boo, we have to go to disk. */
2208 /* read bh, cast, ocfs2_refresh_inode */
2209 status = ocfs2_read_inode_block(inode, bh);
2210 if (status < 0) {
2211 mlog_errno(status);
2212 goto bail_refresh;
2213 }
2214 fe = (struct ocfs2_dinode *) (*bh)->b_data;
2215
2216 /* This is a good chance to make sure we're not
2217 * locking an invalid object. ocfs2_read_inode_block()
2218 * already checked that the inode block is sane.
2219 *
2220 * We bug on a stale inode here because we checked
2221 * above whether it was wiped from disk. The wiping
2222 * node provides a guarantee that we receive that
2223 * message and can mark the inode before dropping any
2224 * locks associated with it. */
2225 mlog_bug_on_msg(inode->i_generation !=
2226 le32_to_cpu(fe->i_generation),
2227 "Invalid dinode %llu disk generation: %u "
2228 "inode->i_generation: %u\n",
2229 (unsigned long long)oi->ip_blkno,
2230 le32_to_cpu(fe->i_generation),
2231 inode->i_generation);
2232 mlog_bug_on_msg(le64_to_cpu(fe->i_dtime) ||
2233 !(fe->i_flags & cpu_to_le32(OCFS2_VALID_FL)),
2234 "Stale dinode %llu dtime: %llu flags: 0x%x\n",
2235 (unsigned long long)oi->ip_blkno,
2236 (unsigned long long)le64_to_cpu(fe->i_dtime),
2237 le32_to_cpu(fe->i_flags));
2238
2239 ocfs2_refresh_inode(inode, fe);
2240 ocfs2_track_lock_refresh(lockres);
2241 }
2242
2243 status = 0;
2244bail_refresh:
2245 ocfs2_complete_lock_res_refresh(lockres, status);
2246bail:
2247 return status;
2248}
2249
2250static int ocfs2_assign_bh(struct inode *inode,
2251 struct buffer_head **ret_bh,
2252 struct buffer_head *passed_bh)
2253{
2254 int status;
2255
2256 if (passed_bh) {
2257 /* Ok, the update went to disk for us, use the
2258 * returned bh. */
2259 *ret_bh = passed_bh;
2260 get_bh(*ret_bh);
2261
2262 return 0;
2263 }
2264
2265 status = ocfs2_read_inode_block(inode, ret_bh);
2266 if (status < 0)
2267 mlog_errno(status);
2268
2269 return status;
2270}
2271
2272/*
2273 * returns < 0 error if the callback will never be called, otherwise
2274 * the result of the lock will be communicated via the callback.
2275 */
2276int ocfs2_inode_lock_full_nested(struct inode *inode,
2277 struct buffer_head **ret_bh,
2278 int ex,
2279 int arg_flags,
2280 int subclass)
2281{
2282 int status, level, acquired;
2283 u32 dlm_flags;
2284 struct ocfs2_lock_res *lockres = NULL;
2285 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
2286 struct buffer_head *local_bh = NULL;
2287
2288 BUG_ON(!inode);
2289
2290 mlog(0, "inode %llu, take %s META lock\n",
2291 (unsigned long long)OCFS2_I(inode)->ip_blkno,
2292 ex ? "EXMODE" : "PRMODE");
2293
2294 status = 0;
2295 acquired = 0;
2296 /* We'll allow faking a readonly metadata lock for
2297 * rodevices. */
2298 if (ocfs2_is_hard_readonly(osb)) {
2299 if (ex)
2300 status = -EROFS;
2301 goto bail;
2302 }
2303
2304 if (ocfs2_mount_local(osb))
2305 goto local;
2306
2307 if (!(arg_flags & OCFS2_META_LOCK_RECOVERY))
2308 ocfs2_wait_for_recovery(osb);
2309
2310 lockres = &OCFS2_I(inode)->ip_inode_lockres;
2311 level = ex ? DLM_LOCK_EX : DLM_LOCK_PR;
2312 dlm_flags = 0;
2313 if (arg_flags & OCFS2_META_LOCK_NOQUEUE)
2314 dlm_flags |= DLM_LKF_NOQUEUE;
2315
2316 status = __ocfs2_cluster_lock(osb, lockres, level, dlm_flags,
2317 arg_flags, subclass, _RET_IP_);
2318 if (status < 0) {
2319 if (status != -EAGAIN && status != -EIOCBRETRY)
2320 mlog_errno(status);
2321 goto bail;
2322 }
2323
2324 /* Notify the error cleanup path to drop the cluster lock. */
2325 acquired = 1;
2326
2327 /* We wait twice because a node may have died while we were in
2328 * the lower dlm layers. The second time though, we've
2329 * committed to owning this lock so we don't allow signals to
2330 * abort the operation. */
2331 if (!(arg_flags & OCFS2_META_LOCK_RECOVERY))
2332 ocfs2_wait_for_recovery(osb);
2333
2334local:
2335 /*
2336 * We only see this flag if we're being called from
2337 * ocfs2_read_locked_inode(). It means we're locking an inode
2338 * which hasn't been populated yet, so clear the refresh flag
2339 * and let the caller handle it.
2340 */
2341 if (inode->i_state & I_NEW) {
2342 status = 0;
2343 if (lockres)
2344 ocfs2_complete_lock_res_refresh(lockres, 0);
2345 goto bail;
2346 }
2347
2348 /* This is fun. The caller may want a bh back, or it may
2349 * not. ocfs2_inode_lock_update definitely wants one in, but
2350 * may or may not read one, depending on what's in the
2351 * LVB. The result of all of this is that we've *only* gone to
2352 * disk if we have to, so the complexity is worthwhile. */
2353 status = ocfs2_inode_lock_update(inode, &local_bh);
2354 if (status < 0) {
2355 if (status != -ENOENT)
2356 mlog_errno(status);
2357 goto bail;
2358 }
2359
2360 if (ret_bh) {
2361 status = ocfs2_assign_bh(inode, ret_bh, local_bh);
2362 if (status < 0) {
2363 mlog_errno(status);
2364 goto bail;
2365 }
2366 }
2367
2368bail:
2369 if (status < 0) {
2370 if (ret_bh && (*ret_bh)) {
2371 brelse(*ret_bh);
2372 *ret_bh = NULL;
2373 }
2374 if (acquired)
2375 ocfs2_inode_unlock(inode, ex);
2376 }
2377
2378 if (local_bh)
2379 brelse(local_bh);
2380
2381 return status;
2382}
2383
2384/*
2385 * This is working around a lock inversion between tasks acquiring DLM
2386 * locks while holding a page lock and the downconvert thread which
2387 * blocks dlm lock acquiry while acquiring page locks.
2388 *
2389 * ** These _with_page variantes are only intended to be called from aop
2390 * methods that hold page locks and return a very specific *positive* error
2391 * code that aop methods pass up to the VFS -- test for errors with != 0. **
2392 *
2393 * The DLM is called such that it returns -EAGAIN if it would have
2394 * blocked waiting for the downconvert thread. In that case we unlock
2395 * our page so the downconvert thread can make progress. Once we've
2396 * done this we have to return AOP_TRUNCATED_PAGE so the aop method
2397 * that called us can bubble that back up into the VFS who will then
2398 * immediately retry the aop call.
2399 *
2400 * We do a blocking lock and immediate unlock before returning, though, so that
2401 * the lock has a great chance of being cached on this node by the time the VFS
2402 * calls back to retry the aop. This has a potential to livelock as nodes
2403 * ping locks back and forth, but that's a risk we're willing to take to avoid
2404 * the lock inversion simply.
2405 */
2406int ocfs2_inode_lock_with_page(struct inode *inode,
2407 struct buffer_head **ret_bh,
2408 int ex,
2409 struct page *page)
2410{
2411 int ret;
2412
2413 ret = ocfs2_inode_lock_full(inode, ret_bh, ex, OCFS2_LOCK_NONBLOCK);
2414 if (ret == -EAGAIN) {
2415 unlock_page(page);
2416 if (ocfs2_inode_lock(inode, ret_bh, ex) == 0)
2417 ocfs2_inode_unlock(inode, ex);
2418 ret = AOP_TRUNCATED_PAGE;
2419 }
2420
2421 return ret;
2422}
2423
2424int ocfs2_inode_lock_atime(struct inode *inode,
2425 struct vfsmount *vfsmnt,
2426 int *level)
2427{
2428 int ret;
2429
2430 ret = ocfs2_inode_lock(inode, NULL, 0);
2431 if (ret < 0) {
2432 mlog_errno(ret);
2433 return ret;
2434 }
2435
2436 /*
2437 * If we should update atime, we will get EX lock,
2438 * otherwise we just get PR lock.
2439 */
2440 if (ocfs2_should_update_atime(inode, vfsmnt)) {
2441 struct buffer_head *bh = NULL;
2442
2443 ocfs2_inode_unlock(inode, 0);
2444 ret = ocfs2_inode_lock(inode, &bh, 1);
2445 if (ret < 0) {
2446 mlog_errno(ret);
2447 return ret;
2448 }
2449 *level = 1;
2450 if (ocfs2_should_update_atime(inode, vfsmnt))
2451 ocfs2_update_inode_atime(inode, bh);
2452 if (bh)
2453 brelse(bh);
2454 } else
2455 *level = 0;
2456
2457 return ret;
2458}
2459
2460void ocfs2_inode_unlock(struct inode *inode,
2461 int ex)
2462{
2463 int level = ex ? DLM_LOCK_EX : DLM_LOCK_PR;
2464 struct ocfs2_lock_res *lockres = &OCFS2_I(inode)->ip_inode_lockres;
2465 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
2466
2467 mlog(0, "inode %llu drop %s META lock\n",
2468 (unsigned long long)OCFS2_I(inode)->ip_blkno,
2469 ex ? "EXMODE" : "PRMODE");
2470
2471 if (!ocfs2_is_hard_readonly(OCFS2_SB(inode->i_sb)) &&
2472 !ocfs2_mount_local(osb))
2473 ocfs2_cluster_unlock(OCFS2_SB(inode->i_sb), lockres, level);
2474}
2475
2476int ocfs2_orphan_scan_lock(struct ocfs2_super *osb, u32 *seqno)
2477{
2478 struct ocfs2_lock_res *lockres;
2479 struct ocfs2_orphan_scan_lvb *lvb;
2480 int status = 0;
2481
2482 if (ocfs2_is_hard_readonly(osb))
2483 return -EROFS;
2484
2485 if (ocfs2_mount_local(osb))
2486 return 0;
2487
2488 lockres = &osb->osb_orphan_scan.os_lockres;
2489 status = ocfs2_cluster_lock(osb, lockres, DLM_LOCK_EX, 0, 0);
2490 if (status < 0)
2491 return status;
2492
2493 lvb = ocfs2_dlm_lvb(&lockres->l_lksb);
2494 if (ocfs2_dlm_lvb_valid(&lockres->l_lksb) &&
2495 lvb->lvb_version == OCFS2_ORPHAN_LVB_VERSION)
2496 *seqno = be32_to_cpu(lvb->lvb_os_seqno);
2497 else
2498 *seqno = osb->osb_orphan_scan.os_seqno + 1;
2499
2500 return status;
2501}
2502
2503void ocfs2_orphan_scan_unlock(struct ocfs2_super *osb, u32 seqno)
2504{
2505 struct ocfs2_lock_res *lockres;
2506 struct ocfs2_orphan_scan_lvb *lvb;
2507
2508 if (!ocfs2_is_hard_readonly(osb) && !ocfs2_mount_local(osb)) {
2509 lockres = &osb->osb_orphan_scan.os_lockres;
2510 lvb = ocfs2_dlm_lvb(&lockres->l_lksb);
2511 lvb->lvb_version = OCFS2_ORPHAN_LVB_VERSION;
2512 lvb->lvb_os_seqno = cpu_to_be32(seqno);
2513 ocfs2_cluster_unlock(osb, lockres, DLM_LOCK_EX);
2514 }
2515}
2516
2517int ocfs2_super_lock(struct ocfs2_super *osb,
2518 int ex)
2519{
2520 int status = 0;
2521 int level = ex ? DLM_LOCK_EX : DLM_LOCK_PR;
2522 struct ocfs2_lock_res *lockres = &osb->osb_super_lockres;
2523
2524 if (ocfs2_is_hard_readonly(osb))
2525 return -EROFS;
2526
2527 if (ocfs2_mount_local(osb))
2528 goto bail;
2529
2530 status = ocfs2_cluster_lock(osb, lockres, level, 0, 0);
2531 if (status < 0) {
2532 mlog_errno(status);
2533 goto bail;
2534 }
2535
2536 /* The super block lock path is really in the best position to
2537 * know when resources covered by the lock need to be
2538 * refreshed, so we do it here. Of course, making sense of
2539 * everything is up to the caller :) */
2540 status = ocfs2_should_refresh_lock_res(lockres);
2541 if (status < 0) {
2542 mlog_errno(status);
2543 goto bail;
2544 }
2545 if (status) {
2546 status = ocfs2_refresh_slot_info(osb);
2547
2548 ocfs2_complete_lock_res_refresh(lockres, status);
2549
2550 if (status < 0)
2551 mlog_errno(status);
2552 ocfs2_track_lock_refresh(lockres);
2553 }
2554bail:
2555 return status;
2556}
2557
2558void ocfs2_super_unlock(struct ocfs2_super *osb,
2559 int ex)
2560{
2561 int level = ex ? DLM_LOCK_EX : DLM_LOCK_PR;
2562 struct ocfs2_lock_res *lockres = &osb->osb_super_lockres;
2563
2564 if (!ocfs2_mount_local(osb))
2565 ocfs2_cluster_unlock(osb, lockres, level);
2566}
2567
2568int ocfs2_rename_lock(struct ocfs2_super *osb)
2569{
2570 int status;
2571 struct ocfs2_lock_res *lockres = &osb->osb_rename_lockres;
2572
2573 if (ocfs2_is_hard_readonly(osb))
2574 return -EROFS;
2575
2576 if (ocfs2_mount_local(osb))
2577 return 0;
2578
2579 status = ocfs2_cluster_lock(osb, lockres, DLM_LOCK_EX, 0, 0);
2580 if (status < 0)
2581 mlog_errno(status);
2582
2583 return status;
2584}
2585
2586void ocfs2_rename_unlock(struct ocfs2_super *osb)
2587{
2588 struct ocfs2_lock_res *lockres = &osb->osb_rename_lockres;
2589
2590 if (!ocfs2_mount_local(osb))
2591 ocfs2_cluster_unlock(osb, lockres, DLM_LOCK_EX);
2592}
2593
2594int ocfs2_nfs_sync_lock(struct ocfs2_super *osb, int ex)
2595{
2596 int status;
2597 struct ocfs2_lock_res *lockres = &osb->osb_nfs_sync_lockres;
2598
2599 if (ocfs2_is_hard_readonly(osb))
2600 return -EROFS;
2601
2602 if (ocfs2_mount_local(osb))
2603 return 0;
2604
2605 status = ocfs2_cluster_lock(osb, lockres, ex ? LKM_EXMODE : LKM_PRMODE,
2606 0, 0);
2607 if (status < 0)
2608 mlog(ML_ERROR, "lock on nfs sync lock failed %d\n", status);
2609
2610 return status;
2611}
2612
2613void ocfs2_nfs_sync_unlock(struct ocfs2_super *osb, int ex)
2614{
2615 struct ocfs2_lock_res *lockres = &osb->osb_nfs_sync_lockres;
2616
2617 if (!ocfs2_mount_local(osb))
2618 ocfs2_cluster_unlock(osb, lockres,
2619 ex ? LKM_EXMODE : LKM_PRMODE);
2620}
2621
2622int ocfs2_dentry_lock(struct dentry *dentry, int ex)
2623{
2624 int ret;
2625 int level = ex ? DLM_LOCK_EX : DLM_LOCK_PR;
2626 struct ocfs2_dentry_lock *dl = dentry->d_fsdata;
2627 struct ocfs2_super *osb = OCFS2_SB(dentry->d_sb);
2628
2629 BUG_ON(!dl);
2630
2631 if (ocfs2_is_hard_readonly(osb))
2632 return -EROFS;
2633
2634 if (ocfs2_mount_local(osb))
2635 return 0;
2636
2637 ret = ocfs2_cluster_lock(osb, &dl->dl_lockres, level, 0, 0);
2638 if (ret < 0)
2639 mlog_errno(ret);
2640
2641 return ret;
2642}
2643
2644void ocfs2_dentry_unlock(struct dentry *dentry, int ex)
2645{
2646 int level = ex ? DLM_LOCK_EX : DLM_LOCK_PR;
2647 struct ocfs2_dentry_lock *dl = dentry->d_fsdata;
2648 struct ocfs2_super *osb = OCFS2_SB(dentry->d_sb);
2649
2650 if (!ocfs2_mount_local(osb))
2651 ocfs2_cluster_unlock(osb, &dl->dl_lockres, level);
2652}
2653
2654/* Reference counting of the dlm debug structure. We want this because
2655 * open references on the debug inodes can live on after a mount, so
2656 * we can't rely on the ocfs2_super to always exist. */
2657static void ocfs2_dlm_debug_free(struct kref *kref)
2658{
2659 struct ocfs2_dlm_debug *dlm_debug;
2660
2661 dlm_debug = container_of(kref, struct ocfs2_dlm_debug, d_refcnt);
2662
2663 kfree(dlm_debug);
2664}
2665
2666void ocfs2_put_dlm_debug(struct ocfs2_dlm_debug *dlm_debug)
2667{
2668 if (dlm_debug)
2669 kref_put(&dlm_debug->d_refcnt, ocfs2_dlm_debug_free);
2670}
2671
2672static void ocfs2_get_dlm_debug(struct ocfs2_dlm_debug *debug)
2673{
2674 kref_get(&debug->d_refcnt);
2675}
2676
2677struct ocfs2_dlm_debug *ocfs2_new_dlm_debug(void)
2678{
2679 struct ocfs2_dlm_debug *dlm_debug;
2680
2681 dlm_debug = kmalloc(sizeof(struct ocfs2_dlm_debug), GFP_KERNEL);
2682 if (!dlm_debug) {
2683 mlog_errno(-ENOMEM);
2684 goto out;
2685 }
2686
2687 kref_init(&dlm_debug->d_refcnt);
2688 INIT_LIST_HEAD(&dlm_debug->d_lockres_tracking);
2689 dlm_debug->d_locking_state = NULL;
2690out:
2691 return dlm_debug;
2692}
2693
2694/* Access to this is arbitrated for us via seq_file->sem. */
2695struct ocfs2_dlm_seq_priv {
2696 struct ocfs2_dlm_debug *p_dlm_debug;
2697 struct ocfs2_lock_res p_iter_res;
2698 struct ocfs2_lock_res p_tmp_res;
2699};
2700
2701static struct ocfs2_lock_res *ocfs2_dlm_next_res(struct ocfs2_lock_res *start,
2702 struct ocfs2_dlm_seq_priv *priv)
2703{
2704 struct ocfs2_lock_res *iter, *ret = NULL;
2705 struct ocfs2_dlm_debug *dlm_debug = priv->p_dlm_debug;
2706
2707 assert_spin_locked(&ocfs2_dlm_tracking_lock);
2708
2709 list_for_each_entry(iter, &start->l_debug_list, l_debug_list) {
2710 /* discover the head of the list */
2711 if (&iter->l_debug_list == &dlm_debug->d_lockres_tracking) {
2712 mlog(0, "End of list found, %p\n", ret);
2713 break;
2714 }
2715
2716 /* We track our "dummy" iteration lockres' by a NULL
2717 * l_ops field. */
2718 if (iter->l_ops != NULL) {
2719 ret = iter;
2720 break;
2721 }
2722 }
2723
2724 return ret;
2725}
2726
2727static void *ocfs2_dlm_seq_start(struct seq_file *m, loff_t *pos)
2728{
2729 struct ocfs2_dlm_seq_priv *priv = m->private;
2730 struct ocfs2_lock_res *iter;
2731
2732 spin_lock(&ocfs2_dlm_tracking_lock);
2733 iter = ocfs2_dlm_next_res(&priv->p_iter_res, priv);
2734 if (iter) {
2735 /* Since lockres' have the lifetime of their container
2736 * (which can be inodes, ocfs2_supers, etc) we want to
2737 * copy this out to a temporary lockres while still
2738 * under the spinlock. Obviously after this we can't
2739 * trust any pointers on the copy returned, but that's
2740 * ok as the information we want isn't typically held
2741 * in them. */
2742 priv->p_tmp_res = *iter;
2743 iter = &priv->p_tmp_res;
2744 }
2745 spin_unlock(&ocfs2_dlm_tracking_lock);
2746
2747 return iter;
2748}
2749
2750static void ocfs2_dlm_seq_stop(struct seq_file *m, void *v)
2751{
2752}
2753
2754static void *ocfs2_dlm_seq_next(struct seq_file *m, void *v, loff_t *pos)
2755{
2756 struct ocfs2_dlm_seq_priv *priv = m->private;
2757 struct ocfs2_lock_res *iter = v;
2758 struct ocfs2_lock_res *dummy = &priv->p_iter_res;
2759
2760 spin_lock(&ocfs2_dlm_tracking_lock);
2761 iter = ocfs2_dlm_next_res(iter, priv);
2762 list_del_init(&dummy->l_debug_list);
2763 if (iter) {
2764 list_add(&dummy->l_debug_list, &iter->l_debug_list);
2765 priv->p_tmp_res = *iter;
2766 iter = &priv->p_tmp_res;
2767 }
2768 spin_unlock(&ocfs2_dlm_tracking_lock);
2769
2770 return iter;
2771}
2772
2773/*
2774 * Version is used by debugfs.ocfs2 to determine the format being used
2775 *
2776 * New in version 2
2777 * - Lock stats printed
2778 * New in version 3
2779 * - Max time in lock stats is in usecs (instead of nsecs)
2780 */
2781#define OCFS2_DLM_DEBUG_STR_VERSION 3
2782static int ocfs2_dlm_seq_show(struct seq_file *m, void *v)
2783{
2784 int i;
2785 char *lvb;
2786 struct ocfs2_lock_res *lockres = v;
2787
2788 if (!lockres)
2789 return -EINVAL;
2790
2791 seq_printf(m, "0x%x\t", OCFS2_DLM_DEBUG_STR_VERSION);
2792
2793 if (lockres->l_type == OCFS2_LOCK_TYPE_DENTRY)
2794 seq_printf(m, "%.*s%08x\t", OCFS2_DENTRY_LOCK_INO_START - 1,
2795 lockres->l_name,
2796 (unsigned int)ocfs2_get_dentry_lock_ino(lockres));
2797 else
2798 seq_printf(m, "%.*s\t", OCFS2_LOCK_ID_MAX_LEN, lockres->l_name);
2799
2800 seq_printf(m, "%d\t"
2801 "0x%lx\t"
2802 "0x%x\t"
2803 "0x%x\t"
2804 "%u\t"
2805 "%u\t"
2806 "%d\t"
2807 "%d\t",
2808 lockres->l_level,
2809 lockres->l_flags,
2810 lockres->l_action,
2811 lockres->l_unlock_action,
2812 lockres->l_ro_holders,
2813 lockres->l_ex_holders,
2814 lockres->l_requested,
2815 lockres->l_blocking);
2816
2817 /* Dump the raw LVB */
2818 lvb = ocfs2_dlm_lvb(&lockres->l_lksb);
2819 for(i = 0; i < DLM_LVB_LEN; i++)
2820 seq_printf(m, "0x%x\t", lvb[i]);
2821
2822#ifdef CONFIG_OCFS2_FS_STATS
2823# define lock_num_prmode(_l) ((_l)->l_lock_prmode.ls_gets)
2824# define lock_num_exmode(_l) ((_l)->l_lock_exmode.ls_gets)
2825# define lock_num_prmode_failed(_l) ((_l)->l_lock_prmode.ls_fail)
2826# define lock_num_exmode_failed(_l) ((_l)->l_lock_exmode.ls_fail)
2827# define lock_total_prmode(_l) ((_l)->l_lock_prmode.ls_total)
2828# define lock_total_exmode(_l) ((_l)->l_lock_exmode.ls_total)
2829# define lock_max_prmode(_l) ((_l)->l_lock_prmode.ls_max)
2830# define lock_max_exmode(_l) ((_l)->l_lock_exmode.ls_max)
2831# define lock_refresh(_l) ((_l)->l_lock_refresh)
2832#else
2833# define lock_num_prmode(_l) (0)
2834# define lock_num_exmode(_l) (0)
2835# define lock_num_prmode_failed(_l) (0)
2836# define lock_num_exmode_failed(_l) (0)
2837# define lock_total_prmode(_l) (0ULL)
2838# define lock_total_exmode(_l) (0ULL)
2839# define lock_max_prmode(_l) (0)
2840# define lock_max_exmode(_l) (0)
2841# define lock_refresh(_l) (0)
2842#endif
2843 /* The following seq_print was added in version 2 of this output */
2844 seq_printf(m, "%u\t"
2845 "%u\t"
2846 "%u\t"
2847 "%u\t"
2848 "%llu\t"
2849 "%llu\t"
2850 "%u\t"
2851 "%u\t"
2852 "%u\t",
2853 lock_num_prmode(lockres),
2854 lock_num_exmode(lockres),
2855 lock_num_prmode_failed(lockres),
2856 lock_num_exmode_failed(lockres),
2857 lock_total_prmode(lockres),
2858 lock_total_exmode(lockres),
2859 lock_max_prmode(lockres),
2860 lock_max_exmode(lockres),
2861 lock_refresh(lockres));
2862
2863 /* End the line */
2864 seq_printf(m, "\n");
2865 return 0;
2866}
2867
2868static const struct seq_operations ocfs2_dlm_seq_ops = {
2869 .start = ocfs2_dlm_seq_start,
2870 .stop = ocfs2_dlm_seq_stop,
2871 .next = ocfs2_dlm_seq_next,
2872 .show = ocfs2_dlm_seq_show,
2873};
2874
2875static int ocfs2_dlm_debug_release(struct inode *inode, struct file *file)
2876{
2877 struct seq_file *seq = file->private_data;
2878 struct ocfs2_dlm_seq_priv *priv = seq->private;
2879 struct ocfs2_lock_res *res = &priv->p_iter_res;
2880
2881 ocfs2_remove_lockres_tracking(res);
2882 ocfs2_put_dlm_debug(priv->p_dlm_debug);
2883 return seq_release_private(inode, file);
2884}
2885
2886static int ocfs2_dlm_debug_open(struct inode *inode, struct file *file)
2887{
2888 int ret;
2889 struct ocfs2_dlm_seq_priv *priv;
2890 struct seq_file *seq;
2891 struct ocfs2_super *osb;
2892
2893 priv = kzalloc(sizeof(struct ocfs2_dlm_seq_priv), GFP_KERNEL);
2894 if (!priv) {
2895 ret = -ENOMEM;
2896 mlog_errno(ret);
2897 goto out;
2898 }
2899 osb = inode->i_private;
2900 ocfs2_get_dlm_debug(osb->osb_dlm_debug);
2901 priv->p_dlm_debug = osb->osb_dlm_debug;
2902 INIT_LIST_HEAD(&priv->p_iter_res.l_debug_list);
2903
2904 ret = seq_open(file, &ocfs2_dlm_seq_ops);
2905 if (ret) {
2906 kfree(priv);
2907 mlog_errno(ret);
2908 goto out;
2909 }
2910
2911 seq = file->private_data;
2912 seq->private = priv;
2913
2914 ocfs2_add_lockres_tracking(&priv->p_iter_res,
2915 priv->p_dlm_debug);
2916
2917out:
2918 return ret;
2919}
2920
2921static const struct file_operations ocfs2_dlm_debug_fops = {
2922 .open = ocfs2_dlm_debug_open,
2923 .release = ocfs2_dlm_debug_release,
2924 .read = seq_read,
2925 .llseek = seq_lseek,
2926};
2927
2928static int ocfs2_dlm_init_debug(struct ocfs2_super *osb)
2929{
2930 int ret = 0;
2931 struct ocfs2_dlm_debug *dlm_debug = osb->osb_dlm_debug;
2932
2933 dlm_debug->d_locking_state = debugfs_create_file("locking_state",
2934 S_IFREG|S_IRUSR,
2935 osb->osb_debug_root,
2936 osb,
2937 &ocfs2_dlm_debug_fops);
2938 if (!dlm_debug->d_locking_state) {
2939 ret = -EINVAL;
2940 mlog(ML_ERROR,
2941 "Unable to create locking state debugfs file.\n");
2942 goto out;
2943 }
2944
2945 ocfs2_get_dlm_debug(dlm_debug);
2946out:
2947 return ret;
2948}
2949
2950static void ocfs2_dlm_shutdown_debug(struct ocfs2_super *osb)
2951{
2952 struct ocfs2_dlm_debug *dlm_debug = osb->osb_dlm_debug;
2953
2954 if (dlm_debug) {
2955 debugfs_remove(dlm_debug->d_locking_state);
2956 ocfs2_put_dlm_debug(dlm_debug);
2957 }
2958}
2959
2960int ocfs2_dlm_init(struct ocfs2_super *osb)
2961{
2962 int status = 0;
2963 struct ocfs2_cluster_connection *conn = NULL;
2964
2965 if (ocfs2_mount_local(osb)) {
2966 osb->node_num = 0;
2967 goto local;
2968 }
2969
2970 status = ocfs2_dlm_init_debug(osb);
2971 if (status < 0) {
2972 mlog_errno(status);
2973 goto bail;
2974 }
2975
2976 /* launch downconvert thread */
2977 osb->dc_task = kthread_run(ocfs2_downconvert_thread, osb, "ocfs2dc");
2978 if (IS_ERR(osb->dc_task)) {
2979 status = PTR_ERR(osb->dc_task);
2980 osb->dc_task = NULL;
2981 mlog_errno(status);
2982 goto bail;
2983 }
2984
2985 /* for now, uuid == domain */
2986 status = ocfs2_cluster_connect(osb->osb_cluster_stack,
2987 osb->uuid_str,
2988 strlen(osb->uuid_str),
2989 &lproto, ocfs2_do_node_down, osb,
2990 &conn);
2991 if (status) {
2992 mlog_errno(status);
2993 goto bail;
2994 }
2995
2996 status = ocfs2_cluster_this_node(&osb->node_num);
2997 if (status < 0) {
2998 mlog_errno(status);
2999 mlog(ML_ERROR,
3000 "could not find this host's node number\n");
3001 ocfs2_cluster_disconnect(conn, 0);
3002 goto bail;
3003 }
3004
3005local:
3006 ocfs2_super_lock_res_init(&osb->osb_super_lockres, osb);
3007 ocfs2_rename_lock_res_init(&osb->osb_rename_lockres, osb);
3008 ocfs2_nfs_sync_lock_res_init(&osb->osb_nfs_sync_lockres, osb);
3009 ocfs2_orphan_scan_lock_res_init(&osb->osb_orphan_scan.os_lockres, osb);
3010
3011 osb->cconn = conn;
3012
3013 status = 0;
3014bail:
3015 if (status < 0) {
3016 ocfs2_dlm_shutdown_debug(osb);
3017 if (osb->dc_task)
3018 kthread_stop(osb->dc_task);
3019 }
3020
3021 return status;
3022}
3023
3024void ocfs2_dlm_shutdown(struct ocfs2_super *osb,
3025 int hangup_pending)
3026{
3027 ocfs2_drop_osb_locks(osb);
3028
3029 /*
3030 * Now that we have dropped all locks and ocfs2_dismount_volume()
3031 * has disabled recovery, the DLM won't be talking to us. It's
3032 * safe to tear things down before disconnecting the cluster.
3033 */
3034
3035 if (osb->dc_task) {
3036 kthread_stop(osb->dc_task);
3037 osb->dc_task = NULL;
3038 }
3039
3040 ocfs2_lock_res_free(&osb->osb_super_lockres);
3041 ocfs2_lock_res_free(&osb->osb_rename_lockres);
3042 ocfs2_lock_res_free(&osb->osb_nfs_sync_lockres);
3043 ocfs2_lock_res_free(&osb->osb_orphan_scan.os_lockres);
3044
3045 ocfs2_cluster_disconnect(osb->cconn, hangup_pending);
3046 osb->cconn = NULL;
3047
3048 ocfs2_dlm_shutdown_debug(osb);
3049}
3050
3051static int ocfs2_drop_lock(struct ocfs2_super *osb,
3052 struct ocfs2_lock_res *lockres)
3053{
3054 int ret;
3055 unsigned long flags;
3056 u32 lkm_flags = 0;
3057
3058 /* We didn't get anywhere near actually using this lockres. */
3059 if (!(lockres->l_flags & OCFS2_LOCK_INITIALIZED))
3060 goto out;
3061
3062 if (lockres->l_ops->flags & LOCK_TYPE_USES_LVB)
3063 lkm_flags |= DLM_LKF_VALBLK;
3064
3065 spin_lock_irqsave(&lockres->l_lock, flags);
3066
3067 mlog_bug_on_msg(!(lockres->l_flags & OCFS2_LOCK_FREEING),
3068 "lockres %s, flags 0x%lx\n",
3069 lockres->l_name, lockres->l_flags);
3070
3071 while (lockres->l_flags & OCFS2_LOCK_BUSY) {
3072 mlog(0, "waiting on busy lock \"%s\": flags = %lx, action = "
3073 "%u, unlock_action = %u\n",
3074 lockres->l_name, lockres->l_flags, lockres->l_action,
3075 lockres->l_unlock_action);
3076
3077 spin_unlock_irqrestore(&lockres->l_lock, flags);
3078
3079 /* XXX: Today we just wait on any busy
3080 * locks... Perhaps we need to cancel converts in the
3081 * future? */
3082 ocfs2_wait_on_busy_lock(lockres);
3083
3084 spin_lock_irqsave(&lockres->l_lock, flags);
3085 }
3086
3087 if (lockres->l_ops->flags & LOCK_TYPE_USES_LVB) {
3088 if (lockres->l_flags & OCFS2_LOCK_ATTACHED &&
3089 lockres->l_level == DLM_LOCK_EX &&
3090 !(lockres->l_flags & OCFS2_LOCK_NEEDS_REFRESH))
3091 lockres->l_ops->set_lvb(lockres);
3092 }
3093
3094 if (lockres->l_flags & OCFS2_LOCK_BUSY)
3095 mlog(ML_ERROR, "destroying busy lock: \"%s\"\n",
3096 lockres->l_name);
3097 if (lockres->l_flags & OCFS2_LOCK_BLOCKED)
3098 mlog(0, "destroying blocked lock: \"%s\"\n", lockres->l_name);
3099
3100 if (!(lockres->l_flags & OCFS2_LOCK_ATTACHED)) {
3101 spin_unlock_irqrestore(&lockres->l_lock, flags);
3102 goto out;
3103 }
3104
3105 lockres_clear_flags(lockres, OCFS2_LOCK_ATTACHED);
3106
3107 /* make sure we never get here while waiting for an ast to
3108 * fire. */
3109 BUG_ON(lockres->l_action != OCFS2_AST_INVALID);
3110
3111 /* is this necessary? */
3112 lockres_or_flags(lockres, OCFS2_LOCK_BUSY);
3113 lockres->l_unlock_action = OCFS2_UNLOCK_DROP_LOCK;
3114 spin_unlock_irqrestore(&lockres->l_lock, flags);
3115
3116 mlog(0, "lock %s\n", lockres->l_name);
3117
3118 ret = ocfs2_dlm_unlock(osb->cconn, &lockres->l_lksb, lkm_flags);
3119 if (ret) {
3120 ocfs2_log_dlm_error("ocfs2_dlm_unlock", ret, lockres);
3121 mlog(ML_ERROR, "lockres flags: %lu\n", lockres->l_flags);
3122 ocfs2_dlm_dump_lksb(&lockres->l_lksb);
3123 BUG();
3124 }
3125 mlog(0, "lock %s, successful return from ocfs2_dlm_unlock\n",
3126 lockres->l_name);
3127
3128 ocfs2_wait_on_busy_lock(lockres);
3129out:
3130 return 0;
3131}
3132
3133/* Mark the lockres as being dropped. It will no longer be
3134 * queued if blocking, but we still may have to wait on it
3135 * being dequeued from the downconvert thread before we can consider
3136 * it safe to drop.
3137 *
3138 * You can *not* attempt to call cluster_lock on this lockres anymore. */
3139void ocfs2_mark_lockres_freeing(struct ocfs2_lock_res *lockres)
3140{
3141 int status;
3142 struct ocfs2_mask_waiter mw;
3143 unsigned long flags;
3144
3145 ocfs2_init_mask_waiter(&mw);
3146
3147 spin_lock_irqsave(&lockres->l_lock, flags);
3148 lockres->l_flags |= OCFS2_LOCK_FREEING;
3149 while (lockres->l_flags & OCFS2_LOCK_QUEUED) {
3150 lockres_add_mask_waiter(lockres, &mw, OCFS2_LOCK_QUEUED, 0);
3151 spin_unlock_irqrestore(&lockres->l_lock, flags);
3152
3153 mlog(0, "Waiting on lockres %s\n", lockres->l_name);
3154
3155 status = ocfs2_wait_for_mask(&mw);
3156 if (status)
3157 mlog_errno(status);
3158
3159 spin_lock_irqsave(&lockres->l_lock, flags);
3160 }
3161 spin_unlock_irqrestore(&lockres->l_lock, flags);
3162}
3163
3164void ocfs2_simple_drop_lockres(struct ocfs2_super *osb,
3165 struct ocfs2_lock_res *lockres)
3166{
3167 int ret;
3168
3169 ocfs2_mark_lockres_freeing(lockres);
3170 ret = ocfs2_drop_lock(osb, lockres);
3171 if (ret)
3172 mlog_errno(ret);
3173}
3174
3175static void ocfs2_drop_osb_locks(struct ocfs2_super *osb)
3176{
3177 ocfs2_simple_drop_lockres(osb, &osb->osb_super_lockres);
3178 ocfs2_simple_drop_lockres(osb, &osb->osb_rename_lockres);
3179 ocfs2_simple_drop_lockres(osb, &osb->osb_nfs_sync_lockres);
3180 ocfs2_simple_drop_lockres(osb, &osb->osb_orphan_scan.os_lockres);
3181}
3182
3183int ocfs2_drop_inode_locks(struct inode *inode)
3184{
3185 int status, err;
3186
3187 /* No need to call ocfs2_mark_lockres_freeing here -
3188 * ocfs2_clear_inode has done it for us. */
3189
3190 err = ocfs2_drop_lock(OCFS2_SB(inode->i_sb),
3191 &OCFS2_I(inode)->ip_open_lockres);
3192 if (err < 0)
3193 mlog_errno(err);
3194
3195 status = err;
3196
3197 err = ocfs2_drop_lock(OCFS2_SB(inode->i_sb),
3198 &OCFS2_I(inode)->ip_inode_lockres);
3199 if (err < 0)
3200 mlog_errno(err);
3201 if (err < 0 && !status)
3202 status = err;
3203
3204 err = ocfs2_drop_lock(OCFS2_SB(inode->i_sb),
3205 &OCFS2_I(inode)->ip_rw_lockres);
3206 if (err < 0)
3207 mlog_errno(err);
3208 if (err < 0 && !status)
3209 status = err;
3210
3211 return status;
3212}
3213
3214static unsigned int ocfs2_prepare_downconvert(struct ocfs2_lock_res *lockres,
3215 int new_level)
3216{
3217 assert_spin_locked(&lockres->l_lock);
3218
3219 BUG_ON(lockres->l_blocking <= DLM_LOCK_NL);
3220
3221 if (lockres->l_level <= new_level) {
3222 mlog(ML_ERROR, "lockres %s, lvl %d <= %d, blcklst %d, mask %d, "
3223 "type %d, flags 0x%lx, hold %d %d, act %d %d, req %d, "
3224 "block %d, pgen %d\n", lockres->l_name, lockres->l_level,
3225 new_level, list_empty(&lockres->l_blocked_list),
3226 list_empty(&lockres->l_mask_waiters), lockres->l_type,
3227 lockres->l_flags, lockres->l_ro_holders,
3228 lockres->l_ex_holders, lockres->l_action,
3229 lockres->l_unlock_action, lockres->l_requested,
3230 lockres->l_blocking, lockres->l_pending_gen);
3231 BUG();
3232 }
3233
3234 mlog(ML_BASTS, "lockres %s, level %d => %d, blocking %d\n",
3235 lockres->l_name, lockres->l_level, new_level, lockres->l_blocking);
3236
3237 lockres->l_action = OCFS2_AST_DOWNCONVERT;
3238 lockres->l_requested = new_level;
3239 lockres_or_flags(lockres, OCFS2_LOCK_BUSY);
3240 return lockres_set_pending(lockres);
3241}
3242
3243static int ocfs2_downconvert_lock(struct ocfs2_super *osb,
3244 struct ocfs2_lock_res *lockres,
3245 int new_level,
3246 int lvb,
3247 unsigned int generation)
3248{
3249 int ret;
3250 u32 dlm_flags = DLM_LKF_CONVERT;
3251
3252 mlog(ML_BASTS, "lockres %s, level %d => %d\n", lockres->l_name,
3253 lockres->l_level, new_level);
3254
3255 if (lvb)
3256 dlm_flags |= DLM_LKF_VALBLK;
3257
3258 ret = ocfs2_dlm_lock(osb->cconn,
3259 new_level,
3260 &lockres->l_lksb,
3261 dlm_flags,
3262 lockres->l_name,
3263 OCFS2_LOCK_ID_MAX_LEN - 1);
3264 lockres_clear_pending(lockres, generation, osb);
3265 if (ret) {
3266 ocfs2_log_dlm_error("ocfs2_dlm_lock", ret, lockres);
3267 ocfs2_recover_from_dlm_error(lockres, 1);
3268 goto bail;
3269 }
3270
3271 ret = 0;
3272bail:
3273 return ret;
3274}
3275
3276/* returns 1 when the caller should unlock and call ocfs2_dlm_unlock */
3277static int ocfs2_prepare_cancel_convert(struct ocfs2_super *osb,
3278 struct ocfs2_lock_res *lockres)
3279{
3280 assert_spin_locked(&lockres->l_lock);
3281
3282 if (lockres->l_unlock_action == OCFS2_UNLOCK_CANCEL_CONVERT) {
3283 /* If we're already trying to cancel a lock conversion
3284 * then just drop the spinlock and allow the caller to
3285 * requeue this lock. */
3286 mlog(ML_BASTS, "lockres %s, skip convert\n", lockres->l_name);
3287 return 0;
3288 }
3289
3290 /* were we in a convert when we got the bast fire? */
3291 BUG_ON(lockres->l_action != OCFS2_AST_CONVERT &&
3292 lockres->l_action != OCFS2_AST_DOWNCONVERT);
3293 /* set things up for the unlockast to know to just
3294 * clear out the ast_action and unset busy, etc. */
3295 lockres->l_unlock_action = OCFS2_UNLOCK_CANCEL_CONVERT;
3296
3297 mlog_bug_on_msg(!(lockres->l_flags & OCFS2_LOCK_BUSY),
3298 "lock %s, invalid flags: 0x%lx\n",
3299 lockres->l_name, lockres->l_flags);
3300
3301 mlog(ML_BASTS, "lockres %s\n", lockres->l_name);
3302
3303 return 1;
3304}
3305
3306static int ocfs2_cancel_convert(struct ocfs2_super *osb,
3307 struct ocfs2_lock_res *lockres)
3308{
3309 int ret;
3310
3311 ret = ocfs2_dlm_unlock(osb->cconn, &lockres->l_lksb,
3312 DLM_LKF_CANCEL);
3313 if (ret) {
3314 ocfs2_log_dlm_error("ocfs2_dlm_unlock", ret, lockres);
3315 ocfs2_recover_from_dlm_error(lockres, 0);
3316 }
3317
3318 mlog(ML_BASTS, "lockres %s\n", lockres->l_name);
3319
3320 return ret;
3321}
3322
3323static int ocfs2_unblock_lock(struct ocfs2_super *osb,
3324 struct ocfs2_lock_res *lockres,
3325 struct ocfs2_unblock_ctl *ctl)
3326{
3327 unsigned long flags;
3328 int blocking;
3329 int new_level;
3330 int level;
3331 int ret = 0;
3332 int set_lvb = 0;
3333 unsigned int gen;
3334
3335 spin_lock_irqsave(&lockres->l_lock, flags);
3336
3337recheck:
3338 /*
3339 * Is it still blocking? If not, we have no more work to do.
3340 */
3341 if (!(lockres->l_flags & OCFS2_LOCK_BLOCKED)) {
3342 BUG_ON(lockres->l_blocking != DLM_LOCK_NL);
3343 spin_unlock_irqrestore(&lockres->l_lock, flags);
3344 ret = 0;
3345 goto leave;
3346 }
3347
3348 if (lockres->l_flags & OCFS2_LOCK_BUSY) {
3349 /* XXX
3350 * This is a *big* race. The OCFS2_LOCK_PENDING flag
3351 * exists entirely for one reason - another thread has set
3352 * OCFS2_LOCK_BUSY, but has *NOT* yet called dlm_lock().
3353 *
3354 * If we do ocfs2_cancel_convert() before the other thread
3355 * calls dlm_lock(), our cancel will do nothing. We will
3356 * get no ast, and we will have no way of knowing the
3357 * cancel failed. Meanwhile, the other thread will call
3358 * into dlm_lock() and wait...forever.
3359 *
3360 * Why forever? Because another node has asked for the
3361 * lock first; that's why we're here in unblock_lock().
3362 *
3363 * The solution is OCFS2_LOCK_PENDING. When PENDING is
3364 * set, we just requeue the unblock. Only when the other
3365 * thread has called dlm_lock() and cleared PENDING will
3366 * we then cancel their request.
3367 *
3368 * All callers of dlm_lock() must set OCFS2_DLM_PENDING
3369 * at the same time they set OCFS2_DLM_BUSY. They must
3370 * clear OCFS2_DLM_PENDING after dlm_lock() returns.
3371 */
3372 if (lockres->l_flags & OCFS2_LOCK_PENDING) {
3373 mlog(ML_BASTS, "lockres %s, ReQ: Pending\n",
3374 lockres->l_name);
3375 goto leave_requeue;
3376 }
3377
3378 ctl->requeue = 1;
3379 ret = ocfs2_prepare_cancel_convert(osb, lockres);
3380 spin_unlock_irqrestore(&lockres->l_lock, flags);
3381 if (ret) {
3382 ret = ocfs2_cancel_convert(osb, lockres);
3383 if (ret < 0)
3384 mlog_errno(ret);
3385 }
3386 goto leave;
3387 }
3388
3389 /*
3390 * This prevents livelocks. OCFS2_LOCK_UPCONVERT_FINISHING flag is
3391 * set when the ast is received for an upconvert just before the
3392 * OCFS2_LOCK_BUSY flag is cleared. Now if the fs received a bast
3393 * on the heels of the ast, we want to delay the downconvert just
3394 * enough to allow the up requestor to do its task. Because this
3395 * lock is in the blocked queue, the lock will be downconverted
3396 * as soon as the requestor is done with the lock.
3397 */
3398 if (lockres->l_flags & OCFS2_LOCK_UPCONVERT_FINISHING)
3399 goto leave_requeue;
3400
3401 /*
3402 * How can we block and yet be at NL? We were trying to upconvert
3403 * from NL and got canceled. The code comes back here, and now
3404 * we notice and clear BLOCKING.
3405 */
3406 if (lockres->l_level == DLM_LOCK_NL) {
3407 BUG_ON(lockres->l_ex_holders || lockres->l_ro_holders);
3408 mlog(ML_BASTS, "lockres %s, Aborting dc\n", lockres->l_name);
3409 lockres->l_blocking = DLM_LOCK_NL;
3410 lockres_clear_flags(lockres, OCFS2_LOCK_BLOCKED);
3411 spin_unlock_irqrestore(&lockres->l_lock, flags);
3412 goto leave;
3413 }
3414
3415 /* if we're blocking an exclusive and we have *any* holders,
3416 * then requeue. */
3417 if ((lockres->l_blocking == DLM_LOCK_EX)
3418 && (lockres->l_ex_holders || lockres->l_ro_holders)) {
3419 mlog(ML_BASTS, "lockres %s, ReQ: EX/PR Holders %u,%u\n",
3420 lockres->l_name, lockres->l_ex_holders,
3421 lockres->l_ro_holders);
3422 goto leave_requeue;
3423 }
3424
3425 /* If it's a PR we're blocking, then only
3426 * requeue if we've got any EX holders */
3427 if (lockres->l_blocking == DLM_LOCK_PR &&
3428 lockres->l_ex_holders) {
3429 mlog(ML_BASTS, "lockres %s, ReQ: EX Holders %u\n",
3430 lockres->l_name, lockres->l_ex_holders);
3431 goto leave_requeue;
3432 }
3433
3434 /*
3435 * Can we get a lock in this state if the holder counts are
3436 * zero? The meta data unblock code used to check this.
3437 */
3438 if ((lockres->l_ops->flags & LOCK_TYPE_REQUIRES_REFRESH)
3439 && (lockres->l_flags & OCFS2_LOCK_REFRESHING)) {
3440 mlog(ML_BASTS, "lockres %s, ReQ: Lock Refreshing\n",
3441 lockres->l_name);
3442 goto leave_requeue;
3443 }
3444
3445 new_level = ocfs2_highest_compat_lock_level(lockres->l_blocking);
3446
3447 if (lockres->l_ops->check_downconvert
3448 && !lockres->l_ops->check_downconvert(lockres, new_level)) {
3449 mlog(ML_BASTS, "lockres %s, ReQ: Checkpointing\n",
3450 lockres->l_name);
3451 goto leave_requeue;
3452 }
3453
3454 /* If we get here, then we know that there are no more
3455 * incompatible holders (and anyone asking for an incompatible
3456 * lock is blocked). We can now downconvert the lock */
3457 if (!lockres->l_ops->downconvert_worker)
3458 goto downconvert;
3459
3460 /* Some lockres types want to do a bit of work before
3461 * downconverting a lock. Allow that here. The worker function
3462 * may sleep, so we save off a copy of what we're blocking as
3463 * it may change while we're not holding the spin lock. */
3464 blocking = lockres->l_blocking;
3465 level = lockres->l_level;
3466 spin_unlock_irqrestore(&lockres->l_lock, flags);
3467
3468 ctl->unblock_action = lockres->l_ops->downconvert_worker(lockres, blocking);
3469
3470 if (ctl->unblock_action == UNBLOCK_STOP_POST) {
3471 mlog(ML_BASTS, "lockres %s, UNBLOCK_STOP_POST\n",
3472 lockres->l_name);
3473 goto leave;
3474 }
3475
3476 spin_lock_irqsave(&lockres->l_lock, flags);
3477 if ((blocking != lockres->l_blocking) || (level != lockres->l_level)) {
3478 /* If this changed underneath us, then we can't drop
3479 * it just yet. */
3480 mlog(ML_BASTS, "lockres %s, block=%d:%d, level=%d:%d, "
3481 "Recheck\n", lockres->l_name, blocking,
3482 lockres->l_blocking, level, lockres->l_level);
3483 goto recheck;
3484 }
3485
3486downconvert:
3487 ctl->requeue = 0;
3488
3489 if (lockres->l_ops->flags & LOCK_TYPE_USES_LVB) {
3490 if (lockres->l_level == DLM_LOCK_EX)
3491 set_lvb = 1;
3492
3493 /*
3494 * We only set the lvb if the lock has been fully
3495 * refreshed - otherwise we risk setting stale
3496 * data. Otherwise, there's no need to actually clear
3497 * out the lvb here as it's value is still valid.
3498 */
3499 if (set_lvb && !(lockres->l_flags & OCFS2_LOCK_NEEDS_REFRESH))
3500 lockres->l_ops->set_lvb(lockres);
3501 }
3502
3503 gen = ocfs2_prepare_downconvert(lockres, new_level);
3504 spin_unlock_irqrestore(&lockres->l_lock, flags);
3505 ret = ocfs2_downconvert_lock(osb, lockres, new_level, set_lvb,
3506 gen);
3507
3508leave:
3509 if (ret)
3510 mlog_errno(ret);
3511 return ret;
3512
3513leave_requeue:
3514 spin_unlock_irqrestore(&lockres->l_lock, flags);
3515 ctl->requeue = 1;
3516
3517 return 0;
3518}
3519
3520static int ocfs2_data_convert_worker(struct ocfs2_lock_res *lockres,
3521 int blocking)
3522{
3523 struct inode *inode;
3524 struct address_space *mapping;
3525 struct ocfs2_inode_info *oi;
3526
3527 inode = ocfs2_lock_res_inode(lockres);
3528 mapping = inode->i_mapping;
3529
3530 if (S_ISDIR(inode->i_mode)) {
3531 oi = OCFS2_I(inode);
3532 oi->ip_dir_lock_gen++;
3533 mlog(0, "generation: %u\n", oi->ip_dir_lock_gen);
3534 goto out;
3535 }
3536
3537 if (!S_ISREG(inode->i_mode))
3538 goto out;
3539
3540 /*
3541 * We need this before the filemap_fdatawrite() so that it can
3542 * transfer the dirty bit from the PTE to the
3543 * page. Unfortunately this means that even for EX->PR
3544 * downconverts, we'll lose our mappings and have to build
3545 * them up again.
3546 */
3547 unmap_mapping_range(mapping, 0, 0, 0);
3548
3549 if (filemap_fdatawrite(mapping)) {
3550 mlog(ML_ERROR, "Could not sync inode %llu for downconvert!",
3551 (unsigned long long)OCFS2_I(inode)->ip_blkno);
3552 }
3553 sync_mapping_buffers(mapping);
3554 if (blocking == DLM_LOCK_EX) {
3555 truncate_inode_pages(mapping, 0);
3556 } else {
3557 /* We only need to wait on the I/O if we're not also
3558 * truncating pages because truncate_inode_pages waits
3559 * for us above. We don't truncate pages if we're
3560 * blocking anything < EXMODE because we want to keep
3561 * them around in that case. */
3562 filemap_fdatawait(mapping);
3563 }
3564
3565out:
3566 return UNBLOCK_CONTINUE;
3567}
3568
3569static int ocfs2_ci_checkpointed(struct ocfs2_caching_info *ci,
3570 struct ocfs2_lock_res *lockres,
3571 int new_level)
3572{
3573 int checkpointed = ocfs2_ci_fully_checkpointed(ci);
3574
3575 BUG_ON(new_level != DLM_LOCK_NL && new_level != DLM_LOCK_PR);
3576 BUG_ON(lockres->l_level != DLM_LOCK_EX && !checkpointed);
3577
3578 if (checkpointed)
3579 return 1;
3580
3581 ocfs2_start_checkpoint(OCFS2_SB(ocfs2_metadata_cache_get_super(ci)));
3582 return 0;
3583}
3584
3585static int ocfs2_check_meta_downconvert(struct ocfs2_lock_res *lockres,
3586 int new_level)
3587{
3588 struct inode *inode = ocfs2_lock_res_inode(lockres);
3589
3590 return ocfs2_ci_checkpointed(INODE_CACHE(inode), lockres, new_level);
3591}
3592
3593static void ocfs2_set_meta_lvb(struct ocfs2_lock_res *lockres)
3594{
3595 struct inode *inode = ocfs2_lock_res_inode(lockres);
3596
3597 __ocfs2_stuff_meta_lvb(inode);
3598}
3599
3600/*
3601 * Does the final reference drop on our dentry lock. Right now this
3602 * happens in the downconvert thread, but we could choose to simplify the
3603 * dlmglue API and push these off to the ocfs2_wq in the future.
3604 */
3605static void ocfs2_dentry_post_unlock(struct ocfs2_super *osb,
3606 struct ocfs2_lock_res *lockres)
3607{
3608 struct ocfs2_dentry_lock *dl = ocfs2_lock_res_dl(lockres);
3609 ocfs2_dentry_lock_put(osb, dl);
3610}
3611
3612/*
3613 * d_delete() matching dentries before the lock downconvert.
3614 *
3615 * At this point, any process waiting to destroy the
3616 * dentry_lock due to last ref count is stopped by the
3617 * OCFS2_LOCK_QUEUED flag.
3618 *
3619 * We have two potential problems
3620 *
3621 * 1) If we do the last reference drop on our dentry_lock (via dput)
3622 * we'll wind up in ocfs2_release_dentry_lock(), waiting on
3623 * the downconvert to finish. Instead we take an elevated
3624 * reference and push the drop until after we've completed our
3625 * unblock processing.
3626 *
3627 * 2) There might be another process with a final reference,
3628 * waiting on us to finish processing. If this is the case, we
3629 * detect it and exit out - there's no more dentries anyway.
3630 */
3631static int ocfs2_dentry_convert_worker(struct ocfs2_lock_res *lockres,
3632 int blocking)
3633{
3634 struct ocfs2_dentry_lock *dl = ocfs2_lock_res_dl(lockres);
3635 struct ocfs2_inode_info *oi = OCFS2_I(dl->dl_inode);
3636 struct dentry *dentry;
3637 unsigned long flags;
3638 int extra_ref = 0;
3639
3640 /*
3641 * This node is blocking another node from getting a read
3642 * lock. This happens when we've renamed within a
3643 * directory. We've forced the other nodes to d_delete(), but
3644 * we never actually dropped our lock because it's still
3645 * valid. The downconvert code will retain a PR for this node,
3646 * so there's no further work to do.
3647 */
3648 if (blocking == DLM_LOCK_PR)
3649 return UNBLOCK_CONTINUE;
3650
3651 /*
3652 * Mark this inode as potentially orphaned. The code in
3653 * ocfs2_delete_inode() will figure out whether it actually
3654 * needs to be freed or not.
3655 */
3656 spin_lock(&oi->ip_lock);
3657 oi->ip_flags |= OCFS2_INODE_MAYBE_ORPHANED;
3658 spin_unlock(&oi->ip_lock);
3659
3660 /*
3661 * Yuck. We need to make sure however that the check of
3662 * OCFS2_LOCK_FREEING and the extra reference are atomic with
3663 * respect to a reference decrement or the setting of that
3664 * flag.
3665 */
3666 spin_lock_irqsave(&lockres->l_lock, flags);
3667 spin_lock(&dentry_attach_lock);
3668 if (!(lockres->l_flags & OCFS2_LOCK_FREEING)
3669 && dl->dl_count) {
3670 dl->dl_count++;
3671 extra_ref = 1;
3672 }
3673 spin_unlock(&dentry_attach_lock);
3674 spin_unlock_irqrestore(&lockres->l_lock, flags);
3675
3676 mlog(0, "extra_ref = %d\n", extra_ref);
3677
3678 /*
3679 * We have a process waiting on us in ocfs2_dentry_iput(),
3680 * which means we can't have any more outstanding
3681 * aliases. There's no need to do any more work.
3682 */
3683 if (!extra_ref)
3684 return UNBLOCK_CONTINUE;
3685
3686 spin_lock(&dentry_attach_lock);
3687 while (1) {
3688 dentry = ocfs2_find_local_alias(dl->dl_inode,
3689 dl->dl_parent_blkno, 1);
3690 if (!dentry)
3691 break;
3692 spin_unlock(&dentry_attach_lock);
3693
3694 mlog(0, "d_delete(%.*s);\n", dentry->d_name.len,
3695 dentry->d_name.name);
3696
3697 /*
3698 * The following dcache calls may do an
3699 * iput(). Normally we don't want that from the
3700 * downconverting thread, but in this case it's ok
3701 * because the requesting node already has an
3702 * exclusive lock on the inode, so it can't be queued
3703 * for a downconvert.
3704 */
3705 d_delete(dentry);
3706 dput(dentry);
3707
3708 spin_lock(&dentry_attach_lock);
3709 }
3710 spin_unlock(&dentry_attach_lock);
3711
3712 /*
3713 * If we are the last holder of this dentry lock, there is no
3714 * reason to downconvert so skip straight to the unlock.
3715 */
3716 if (dl->dl_count == 1)
3717 return UNBLOCK_STOP_POST;
3718
3719 return UNBLOCK_CONTINUE_POST;
3720}
3721
3722static int ocfs2_check_refcount_downconvert(struct ocfs2_lock_res *lockres,
3723 int new_level)
3724{
3725 struct ocfs2_refcount_tree *tree =
3726 ocfs2_lock_res_refcount_tree(lockres);
3727
3728 return ocfs2_ci_checkpointed(&tree->rf_ci, lockres, new_level);
3729}
3730
3731static int ocfs2_refcount_convert_worker(struct ocfs2_lock_res *lockres,
3732 int blocking)
3733{
3734 struct ocfs2_refcount_tree *tree =
3735 ocfs2_lock_res_refcount_tree(lockres);
3736
3737 ocfs2_metadata_cache_purge(&tree->rf_ci);
3738
3739 return UNBLOCK_CONTINUE;
3740}
3741
3742static void ocfs2_set_qinfo_lvb(struct ocfs2_lock_res *lockres)
3743{
3744 struct ocfs2_qinfo_lvb *lvb;
3745 struct ocfs2_mem_dqinfo *oinfo = ocfs2_lock_res_qinfo(lockres);
3746 struct mem_dqinfo *info = sb_dqinfo(oinfo->dqi_gi.dqi_sb,
3747 oinfo->dqi_gi.dqi_type);
3748
3749 lvb = ocfs2_dlm_lvb(&lockres->l_lksb);
3750 lvb->lvb_version = OCFS2_QINFO_LVB_VERSION;
3751 lvb->lvb_bgrace = cpu_to_be32(info->dqi_bgrace);
3752 lvb->lvb_igrace = cpu_to_be32(info->dqi_igrace);
3753 lvb->lvb_syncms = cpu_to_be32(oinfo->dqi_syncms);
3754 lvb->lvb_blocks = cpu_to_be32(oinfo->dqi_gi.dqi_blocks);
3755 lvb->lvb_free_blk = cpu_to_be32(oinfo->dqi_gi.dqi_free_blk);
3756 lvb->lvb_free_entry = cpu_to_be32(oinfo->dqi_gi.dqi_free_entry);
3757}
3758
3759void ocfs2_qinfo_unlock(struct ocfs2_mem_dqinfo *oinfo, int ex)
3760{
3761 struct ocfs2_lock_res *lockres = &oinfo->dqi_gqlock;
3762 struct ocfs2_super *osb = OCFS2_SB(oinfo->dqi_gi.dqi_sb);
3763 int level = ex ? DLM_LOCK_EX : DLM_LOCK_PR;
3764
3765 if (!ocfs2_is_hard_readonly(osb) && !ocfs2_mount_local(osb))
3766 ocfs2_cluster_unlock(osb, lockres, level);
3767}
3768
3769static int ocfs2_refresh_qinfo(struct ocfs2_mem_dqinfo *oinfo)
3770{
3771 struct mem_dqinfo *info = sb_dqinfo(oinfo->dqi_gi.dqi_sb,
3772 oinfo->dqi_gi.dqi_type);
3773 struct ocfs2_lock_res *lockres = &oinfo->dqi_gqlock;
3774 struct ocfs2_qinfo_lvb *lvb = ocfs2_dlm_lvb(&lockres->l_lksb);
3775 struct buffer_head *bh = NULL;
3776 struct ocfs2_global_disk_dqinfo *gdinfo;
3777 int status = 0;
3778
3779 if (ocfs2_dlm_lvb_valid(&lockres->l_lksb) &&
3780 lvb->lvb_version == OCFS2_QINFO_LVB_VERSION) {
3781 info->dqi_bgrace = be32_to_cpu(lvb->lvb_bgrace);
3782 info->dqi_igrace = be32_to_cpu(lvb->lvb_igrace);
3783 oinfo->dqi_syncms = be32_to_cpu(lvb->lvb_syncms);
3784 oinfo->dqi_gi.dqi_blocks = be32_to_cpu(lvb->lvb_blocks);
3785 oinfo->dqi_gi.dqi_free_blk = be32_to_cpu(lvb->lvb_free_blk);
3786 oinfo->dqi_gi.dqi_free_entry =
3787 be32_to_cpu(lvb->lvb_free_entry);
3788 } else {
3789 status = ocfs2_read_quota_phys_block(oinfo->dqi_gqinode,
3790 oinfo->dqi_giblk, &bh);
3791 if (status) {
3792 mlog_errno(status);
3793 goto bail;
3794 }
3795 gdinfo = (struct ocfs2_global_disk_dqinfo *)
3796 (bh->b_data + OCFS2_GLOBAL_INFO_OFF);
3797 info->dqi_bgrace = le32_to_cpu(gdinfo->dqi_bgrace);
3798 info->dqi_igrace = le32_to_cpu(gdinfo->dqi_igrace);
3799 oinfo->dqi_syncms = le32_to_cpu(gdinfo->dqi_syncms);
3800 oinfo->dqi_gi.dqi_blocks = le32_to_cpu(gdinfo->dqi_blocks);
3801 oinfo->dqi_gi.dqi_free_blk = le32_to_cpu(gdinfo->dqi_free_blk);
3802 oinfo->dqi_gi.dqi_free_entry =
3803 le32_to_cpu(gdinfo->dqi_free_entry);
3804 brelse(bh);
3805 ocfs2_track_lock_refresh(lockres);
3806 }
3807
3808bail:
3809 return status;
3810}
3811
3812/* Lock quota info, this function expects at least shared lock on the quota file
3813 * so that we can safely refresh quota info from disk. */
3814int ocfs2_qinfo_lock(struct ocfs2_mem_dqinfo *oinfo, int ex)
3815{
3816 struct ocfs2_lock_res *lockres = &oinfo->dqi_gqlock;
3817 struct ocfs2_super *osb = OCFS2_SB(oinfo->dqi_gi.dqi_sb);
3818 int level = ex ? DLM_LOCK_EX : DLM_LOCK_PR;
3819 int status = 0;
3820
3821 /* On RO devices, locking really isn't needed... */
3822 if (ocfs2_is_hard_readonly(osb)) {
3823 if (ex)
3824 status = -EROFS;
3825 goto bail;
3826 }
3827 if (ocfs2_mount_local(osb))
3828 goto bail;
3829
3830 status = ocfs2_cluster_lock(osb, lockres, level, 0, 0);
3831 if (status < 0) {
3832 mlog_errno(status);
3833 goto bail;
3834 }
3835 if (!ocfs2_should_refresh_lock_res(lockres))
3836 goto bail;
3837 /* OK, we have the lock but we need to refresh the quota info */
3838 status = ocfs2_refresh_qinfo(oinfo);
3839 if (status)
3840 ocfs2_qinfo_unlock(oinfo, ex);
3841 ocfs2_complete_lock_res_refresh(lockres, status);
3842bail:
3843 return status;
3844}
3845
3846int ocfs2_refcount_lock(struct ocfs2_refcount_tree *ref_tree, int ex)
3847{
3848 int status;
3849 int level = ex ? DLM_LOCK_EX : DLM_LOCK_PR;
3850 struct ocfs2_lock_res *lockres = &ref_tree->rf_lockres;
3851 struct ocfs2_super *osb = lockres->l_priv;
3852
3853
3854 if (ocfs2_is_hard_readonly(osb))
3855 return -EROFS;
3856
3857 if (ocfs2_mount_local(osb))
3858 return 0;
3859
3860 status = ocfs2_cluster_lock(osb, lockres, level, 0, 0);
3861 if (status < 0)
3862 mlog_errno(status);
3863
3864 return status;
3865}
3866
3867void ocfs2_refcount_unlock(struct ocfs2_refcount_tree *ref_tree, int ex)
3868{
3869 int level = ex ? DLM_LOCK_EX : DLM_LOCK_PR;
3870 struct ocfs2_lock_res *lockres = &ref_tree->rf_lockres;
3871 struct ocfs2_super *osb = lockres->l_priv;
3872
3873 if (!ocfs2_mount_local(osb))
3874 ocfs2_cluster_unlock(osb, lockres, level);
3875}
3876
3877static void ocfs2_process_blocked_lock(struct ocfs2_super *osb,
3878 struct ocfs2_lock_res *lockres)
3879{
3880 int status;
3881 struct ocfs2_unblock_ctl ctl = {0, 0,};
3882 unsigned long flags;
3883
3884 /* Our reference to the lockres in this function can be
3885 * considered valid until we remove the OCFS2_LOCK_QUEUED
3886 * flag. */
3887
3888 BUG_ON(!lockres);
3889 BUG_ON(!lockres->l_ops);
3890
3891 mlog(ML_BASTS, "lockres %s blocked\n", lockres->l_name);
3892
3893 /* Detect whether a lock has been marked as going away while
3894 * the downconvert thread was processing other things. A lock can
3895 * still be marked with OCFS2_LOCK_FREEING after this check,
3896 * but short circuiting here will still save us some
3897 * performance. */
3898 spin_lock_irqsave(&lockres->l_lock, flags);
3899 if (lockres->l_flags & OCFS2_LOCK_FREEING)
3900 goto unqueue;
3901 spin_unlock_irqrestore(&lockres->l_lock, flags);
3902
3903 status = ocfs2_unblock_lock(osb, lockres, &ctl);
3904 if (status < 0)
3905 mlog_errno(status);
3906
3907 spin_lock_irqsave(&lockres->l_lock, flags);
3908unqueue:
3909 if (lockres->l_flags & OCFS2_LOCK_FREEING || !ctl.requeue) {
3910 lockres_clear_flags(lockres, OCFS2_LOCK_QUEUED);
3911 } else
3912 ocfs2_schedule_blocked_lock(osb, lockres);
3913
3914 mlog(ML_BASTS, "lockres %s, requeue = %s.\n", lockres->l_name,
3915 ctl.requeue ? "yes" : "no");
3916 spin_unlock_irqrestore(&lockres->l_lock, flags);
3917
3918 if (ctl.unblock_action != UNBLOCK_CONTINUE
3919 && lockres->l_ops->post_unlock)
3920 lockres->l_ops->post_unlock(osb, lockres);
3921}
3922
3923static void ocfs2_schedule_blocked_lock(struct ocfs2_super *osb,
3924 struct ocfs2_lock_res *lockres)
3925{
3926 assert_spin_locked(&lockres->l_lock);
3927
3928 if (lockres->l_flags & OCFS2_LOCK_FREEING) {
3929 /* Do not schedule a lock for downconvert when it's on
3930 * the way to destruction - any nodes wanting access
3931 * to the resource will get it soon. */
3932 mlog(ML_BASTS, "lockres %s won't be scheduled: flags 0x%lx\n",
3933 lockres->l_name, lockres->l_flags);
3934 return;
3935 }
3936
3937 lockres_or_flags(lockres, OCFS2_LOCK_QUEUED);
3938
3939 spin_lock(&osb->dc_task_lock);
3940 if (list_empty(&lockres->l_blocked_list)) {
3941 list_add_tail(&lockres->l_blocked_list,
3942 &osb->blocked_lock_list);
3943 osb->blocked_lock_count++;
3944 }
3945 spin_unlock(&osb->dc_task_lock);
3946}
3947
3948static void ocfs2_downconvert_thread_do_work(struct ocfs2_super *osb)
3949{
3950 unsigned long processed;
3951 struct ocfs2_lock_res *lockres;
3952
3953 spin_lock(&osb->dc_task_lock);
3954 /* grab this early so we know to try again if a state change and
3955 * wake happens part-way through our work */
3956 osb->dc_work_sequence = osb->dc_wake_sequence;
3957
3958 processed = osb->blocked_lock_count;
3959 while (processed) {
3960 BUG_ON(list_empty(&osb->blocked_lock_list));
3961
3962 lockres = list_entry(osb->blocked_lock_list.next,
3963 struct ocfs2_lock_res, l_blocked_list);
3964 list_del_init(&lockres->l_blocked_list);
3965 osb->blocked_lock_count--;
3966 spin_unlock(&osb->dc_task_lock);
3967
3968 BUG_ON(!processed);
3969 processed--;
3970
3971 ocfs2_process_blocked_lock(osb, lockres);
3972
3973 spin_lock(&osb->dc_task_lock);
3974 }
3975 spin_unlock(&osb->dc_task_lock);
3976}
3977
3978static int ocfs2_downconvert_thread_lists_empty(struct ocfs2_super *osb)
3979{
3980 int empty = 0;
3981
3982 spin_lock(&osb->dc_task_lock);
3983 if (list_empty(&osb->blocked_lock_list))
3984 empty = 1;
3985
3986 spin_unlock(&osb->dc_task_lock);
3987 return empty;
3988}
3989
3990static int ocfs2_downconvert_thread_should_wake(struct ocfs2_super *osb)
3991{
3992 int should_wake = 0;
3993
3994 spin_lock(&osb->dc_task_lock);
3995 if (osb->dc_work_sequence != osb->dc_wake_sequence)
3996 should_wake = 1;
3997 spin_unlock(&osb->dc_task_lock);
3998
3999 return should_wake;
4000}
4001
4002static int ocfs2_downconvert_thread(void *arg)
4003{
4004 int status = 0;
4005 struct ocfs2_super *osb = arg;
4006
4007 /* only quit once we've been asked to stop and there is no more
4008 * work available */
4009 while (!(kthread_should_stop() &&
4010 ocfs2_downconvert_thread_lists_empty(osb))) {
4011
4012 wait_event_interruptible(osb->dc_event,
4013 ocfs2_downconvert_thread_should_wake(osb) ||
4014 kthread_should_stop());
4015
4016 mlog(0, "downconvert_thread: awoken\n");
4017
4018 ocfs2_downconvert_thread_do_work(osb);
4019 }
4020
4021 osb->dc_task = NULL;
4022 return status;
4023}
4024
4025void ocfs2_wake_downconvert_thread(struct ocfs2_super *osb)
4026{
4027 spin_lock(&osb->dc_task_lock);
4028 /* make sure the voting thread gets a swipe at whatever changes
4029 * the caller may have made to the voting state */
4030 osb->dc_wake_sequence++;
4031 spin_unlock(&osb->dc_task_lock);
4032 wake_up(&osb->dc_event);
4033}