GNU Linux-libre 4.9.308-gnu1
[releases.git] / fs / ocfs2 / journal.c
1 /* -*- mode: c; c-basic-offset: 8; -*-
2  * vim: noexpandtab sw=8 ts=8 sts=0:
3  *
4  * journal.c
5  *
6  * Defines functions of journalling api
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/fs.h>
27 #include <linux/types.h>
28 #include <linux/slab.h>
29 #include <linux/highmem.h>
30 #include <linux/kthread.h>
31 #include <linux/time.h>
32 #include <linux/random.h>
33 #include <linux/delay.h>
34
35 #include <cluster/masklog.h>
36
37 #include "ocfs2.h"
38
39 #include "alloc.h"
40 #include "blockcheck.h"
41 #include "dir.h"
42 #include "dlmglue.h"
43 #include "extent_map.h"
44 #include "heartbeat.h"
45 #include "inode.h"
46 #include "journal.h"
47 #include "localalloc.h"
48 #include "slot_map.h"
49 #include "super.h"
50 #include "sysfile.h"
51 #include "uptodate.h"
52 #include "quota.h"
53 #include "file.h"
54 #include "namei.h"
55
56 #include "buffer_head_io.h"
57 #include "ocfs2_trace.h"
58
59 DEFINE_SPINLOCK(trans_inc_lock);
60
61 #define ORPHAN_SCAN_SCHEDULE_TIMEOUT 300000
62
63 static int ocfs2_force_read_journal(struct inode *inode);
64 static int ocfs2_recover_node(struct ocfs2_super *osb,
65                               int node_num, int slot_num);
66 static int __ocfs2_recovery_thread(void *arg);
67 static int ocfs2_commit_cache(struct ocfs2_super *osb);
68 static int __ocfs2_wait_on_mount(struct ocfs2_super *osb, int quota);
69 static int ocfs2_journal_toggle_dirty(struct ocfs2_super *osb,
70                                       int dirty, int replayed);
71 static int ocfs2_trylock_journal(struct ocfs2_super *osb,
72                                  int slot_num);
73 static int ocfs2_recover_orphans(struct ocfs2_super *osb,
74                                  int slot,
75                                  enum ocfs2_orphan_reco_type orphan_reco_type);
76 static int ocfs2_commit_thread(void *arg);
77 static void ocfs2_queue_recovery_completion(struct ocfs2_journal *journal,
78                                             int slot_num,
79                                             struct ocfs2_dinode *la_dinode,
80                                             struct ocfs2_dinode *tl_dinode,
81                                             struct ocfs2_quota_recovery *qrec,
82                                             enum ocfs2_orphan_reco_type orphan_reco_type);
83
84 static inline int ocfs2_wait_on_mount(struct ocfs2_super *osb)
85 {
86         return __ocfs2_wait_on_mount(osb, 0);
87 }
88
89 static inline int ocfs2_wait_on_quotas(struct ocfs2_super *osb)
90 {
91         return __ocfs2_wait_on_mount(osb, 1);
92 }
93
94 /*
95  * This replay_map is to track online/offline slots, so we could recover
96  * offline slots during recovery and mount
97  */
98
99 enum ocfs2_replay_state {
100         REPLAY_UNNEEDED = 0,    /* Replay is not needed, so ignore this map */
101         REPLAY_NEEDED,          /* Replay slots marked in rm_replay_slots */
102         REPLAY_DONE             /* Replay was already queued */
103 };
104
105 struct ocfs2_replay_map {
106         unsigned int rm_slots;
107         enum ocfs2_replay_state rm_state;
108         unsigned char rm_replay_slots[0];
109 };
110
111 static void ocfs2_replay_map_set_state(struct ocfs2_super *osb, int state)
112 {
113         if (!osb->replay_map)
114                 return;
115
116         /* If we've already queued the replay, we don't have any more to do */
117         if (osb->replay_map->rm_state == REPLAY_DONE)
118                 return;
119
120         osb->replay_map->rm_state = state;
121 }
122
123 int ocfs2_compute_replay_slots(struct ocfs2_super *osb)
124 {
125         struct ocfs2_replay_map *replay_map;
126         int i, node_num;
127
128         /* If replay map is already set, we don't do it again */
129         if (osb->replay_map)
130                 return 0;
131
132         replay_map = kzalloc(sizeof(struct ocfs2_replay_map) +
133                              (osb->max_slots * sizeof(char)), GFP_KERNEL);
134
135         if (!replay_map) {
136                 mlog_errno(-ENOMEM);
137                 return -ENOMEM;
138         }
139
140         spin_lock(&osb->osb_lock);
141
142         replay_map->rm_slots = osb->max_slots;
143         replay_map->rm_state = REPLAY_UNNEEDED;
144
145         /* set rm_replay_slots for offline slot(s) */
146         for (i = 0; i < replay_map->rm_slots; i++) {
147                 if (ocfs2_slot_to_node_num_locked(osb, i, &node_num) == -ENOENT)
148                         replay_map->rm_replay_slots[i] = 1;
149         }
150
151         osb->replay_map = replay_map;
152         spin_unlock(&osb->osb_lock);
153         return 0;
154 }
155
156 static void ocfs2_queue_replay_slots(struct ocfs2_super *osb,
157                 enum ocfs2_orphan_reco_type orphan_reco_type)
158 {
159         struct ocfs2_replay_map *replay_map = osb->replay_map;
160         int i;
161
162         if (!replay_map)
163                 return;
164
165         if (replay_map->rm_state != REPLAY_NEEDED)
166                 return;
167
168         for (i = 0; i < replay_map->rm_slots; i++)
169                 if (replay_map->rm_replay_slots[i])
170                         ocfs2_queue_recovery_completion(osb->journal, i, NULL,
171                                                         NULL, NULL,
172                                                         orphan_reco_type);
173         replay_map->rm_state = REPLAY_DONE;
174 }
175
176 static void ocfs2_free_replay_slots(struct ocfs2_super *osb)
177 {
178         struct ocfs2_replay_map *replay_map = osb->replay_map;
179
180         if (!osb->replay_map)
181                 return;
182
183         kfree(replay_map);
184         osb->replay_map = NULL;
185 }
186
187 int ocfs2_recovery_init(struct ocfs2_super *osb)
188 {
189         struct ocfs2_recovery_map *rm;
190
191         mutex_init(&osb->recovery_lock);
192         osb->disable_recovery = 0;
193         osb->recovery_thread_task = NULL;
194         init_waitqueue_head(&osb->recovery_event);
195
196         rm = kzalloc(sizeof(struct ocfs2_recovery_map) +
197                      osb->max_slots * sizeof(unsigned int),
198                      GFP_KERNEL);
199         if (!rm) {
200                 mlog_errno(-ENOMEM);
201                 return -ENOMEM;
202         }
203
204         rm->rm_entries = (unsigned int *)((char *)rm +
205                                           sizeof(struct ocfs2_recovery_map));
206         osb->recovery_map = rm;
207
208         return 0;
209 }
210
211 /* we can't grab the goofy sem lock from inside wait_event, so we use
212  * memory barriers to make sure that we'll see the null task before
213  * being woken up */
214 static int ocfs2_recovery_thread_running(struct ocfs2_super *osb)
215 {
216         mb();
217         return osb->recovery_thread_task != NULL;
218 }
219
220 void ocfs2_recovery_exit(struct ocfs2_super *osb)
221 {
222         struct ocfs2_recovery_map *rm;
223
224         /* disable any new recovery threads and wait for any currently
225          * running ones to exit. Do this before setting the vol_state. */
226         mutex_lock(&osb->recovery_lock);
227         osb->disable_recovery = 1;
228         mutex_unlock(&osb->recovery_lock);
229         wait_event(osb->recovery_event, !ocfs2_recovery_thread_running(osb));
230
231         /* At this point, we know that no more recovery threads can be
232          * launched, so wait for any recovery completion work to
233          * complete. */
234         if (osb->ocfs2_wq)
235                 flush_workqueue(osb->ocfs2_wq);
236
237         /*
238          * Now that recovery is shut down, and the osb is about to be
239          * freed,  the osb_lock is not taken here.
240          */
241         rm = osb->recovery_map;
242         /* XXX: Should we bug if there are dirty entries? */
243
244         kfree(rm);
245 }
246
247 static int __ocfs2_recovery_map_test(struct ocfs2_super *osb,
248                                      unsigned int node_num)
249 {
250         int i;
251         struct ocfs2_recovery_map *rm = osb->recovery_map;
252
253         assert_spin_locked(&osb->osb_lock);
254
255         for (i = 0; i < rm->rm_used; i++) {
256                 if (rm->rm_entries[i] == node_num)
257                         return 1;
258         }
259
260         return 0;
261 }
262
263 /* Behaves like test-and-set.  Returns the previous value */
264 static int ocfs2_recovery_map_set(struct ocfs2_super *osb,
265                                   unsigned int node_num)
266 {
267         struct ocfs2_recovery_map *rm = osb->recovery_map;
268
269         spin_lock(&osb->osb_lock);
270         if (__ocfs2_recovery_map_test(osb, node_num)) {
271                 spin_unlock(&osb->osb_lock);
272                 return 1;
273         }
274
275         /* XXX: Can this be exploited? Not from o2dlm... */
276         BUG_ON(rm->rm_used >= osb->max_slots);
277
278         rm->rm_entries[rm->rm_used] = node_num;
279         rm->rm_used++;
280         spin_unlock(&osb->osb_lock);
281
282         return 0;
283 }
284
285 static void ocfs2_recovery_map_clear(struct ocfs2_super *osb,
286                                      unsigned int node_num)
287 {
288         int i;
289         struct ocfs2_recovery_map *rm = osb->recovery_map;
290
291         spin_lock(&osb->osb_lock);
292
293         for (i = 0; i < rm->rm_used; i++) {
294                 if (rm->rm_entries[i] == node_num)
295                         break;
296         }
297
298         if (i < rm->rm_used) {
299                 /* XXX: be careful with the pointer math */
300                 memmove(&(rm->rm_entries[i]), &(rm->rm_entries[i + 1]),
301                         (rm->rm_used - i - 1) * sizeof(unsigned int));
302                 rm->rm_used--;
303         }
304
305         spin_unlock(&osb->osb_lock);
306 }
307
308 static int ocfs2_commit_cache(struct ocfs2_super *osb)
309 {
310         int status = 0;
311         unsigned int flushed;
312         struct ocfs2_journal *journal = NULL;
313
314         journal = osb->journal;
315
316         /* Flush all pending commits and checkpoint the journal. */
317         down_write(&journal->j_trans_barrier);
318
319         flushed = atomic_read(&journal->j_num_trans);
320         trace_ocfs2_commit_cache_begin(flushed);
321         if (flushed == 0) {
322                 up_write(&journal->j_trans_barrier);
323                 goto finally;
324         }
325
326         jbd2_journal_lock_updates(journal->j_journal);
327         status = jbd2_journal_flush(journal->j_journal);
328         jbd2_journal_unlock_updates(journal->j_journal);
329         if (status < 0) {
330                 up_write(&journal->j_trans_barrier);
331                 mlog_errno(status);
332                 goto finally;
333         }
334
335         ocfs2_inc_trans_id(journal);
336
337         flushed = atomic_read(&journal->j_num_trans);
338         atomic_set(&journal->j_num_trans, 0);
339         up_write(&journal->j_trans_barrier);
340
341         trace_ocfs2_commit_cache_end(journal->j_trans_id, flushed);
342
343         ocfs2_wake_downconvert_thread(osb);
344         wake_up(&journal->j_checkpointed);
345 finally:
346         return status;
347 }
348
349 handle_t *ocfs2_start_trans(struct ocfs2_super *osb, int max_buffs)
350 {
351         journal_t *journal = osb->journal->j_journal;
352         handle_t *handle;
353
354         BUG_ON(!osb || !osb->journal->j_journal);
355
356         if (ocfs2_is_hard_readonly(osb))
357                 return ERR_PTR(-EROFS);
358
359         BUG_ON(osb->journal->j_state == OCFS2_JOURNAL_FREE);
360         BUG_ON(max_buffs <= 0);
361
362         /* Nested transaction? Just return the handle... */
363         if (journal_current_handle())
364                 return jbd2_journal_start(journal, max_buffs);
365
366         sb_start_intwrite(osb->sb);
367
368         down_read(&osb->journal->j_trans_barrier);
369
370         handle = jbd2_journal_start(journal, max_buffs);
371         if (IS_ERR(handle)) {
372                 up_read(&osb->journal->j_trans_barrier);
373                 sb_end_intwrite(osb->sb);
374
375                 mlog_errno(PTR_ERR(handle));
376
377                 if (is_journal_aborted(journal)) {
378                         ocfs2_abort(osb->sb, "Detected aborted journal\n");
379                         handle = ERR_PTR(-EROFS);
380                 }
381         } else {
382                 if (!ocfs2_mount_local(osb))
383                         atomic_inc(&(osb->journal->j_num_trans));
384         }
385
386         return handle;
387 }
388
389 int ocfs2_commit_trans(struct ocfs2_super *osb,
390                        handle_t *handle)
391 {
392         int ret, nested;
393         struct ocfs2_journal *journal = osb->journal;
394
395         BUG_ON(!handle);
396
397         nested = handle->h_ref > 1;
398         ret = jbd2_journal_stop(handle);
399         if (ret < 0)
400                 mlog_errno(ret);
401
402         if (!nested) {
403                 up_read(&journal->j_trans_barrier);
404                 sb_end_intwrite(osb->sb);
405         }
406
407         return ret;
408 }
409
410 /*
411  * 'nblocks' is what you want to add to the current transaction.
412  *
413  * This might call jbd2_journal_restart() which will commit dirty buffers
414  * and then restart the transaction. Before calling
415  * ocfs2_extend_trans(), any changed blocks should have been
416  * dirtied. After calling it, all blocks which need to be changed must
417  * go through another set of journal_access/journal_dirty calls.
418  *
419  * WARNING: This will not release any semaphores or disk locks taken
420  * during the transaction, so make sure they were taken *before*
421  * start_trans or we'll have ordering deadlocks.
422  *
423  * WARNING2: Note that we do *not* drop j_trans_barrier here. This is
424  * good because transaction ids haven't yet been recorded on the
425  * cluster locks associated with this handle.
426  */
427 int ocfs2_extend_trans(handle_t *handle, int nblocks)
428 {
429         int status, old_nblocks;
430
431         BUG_ON(!handle);
432         BUG_ON(nblocks < 0);
433
434         if (!nblocks)
435                 return 0;
436
437         old_nblocks = handle->h_buffer_credits;
438
439         trace_ocfs2_extend_trans(old_nblocks, nblocks);
440
441 #ifdef CONFIG_OCFS2_DEBUG_FS
442         status = 1;
443 #else
444         status = jbd2_journal_extend(handle, nblocks);
445         if (status < 0) {
446                 mlog_errno(status);
447                 goto bail;
448         }
449 #endif
450
451         if (status > 0) {
452                 trace_ocfs2_extend_trans_restart(old_nblocks + nblocks);
453                 status = jbd2_journal_restart(handle,
454                                               old_nblocks + nblocks);
455                 if (status < 0) {
456                         mlog_errno(status);
457                         goto bail;
458                 }
459         }
460
461         status = 0;
462 bail:
463         return status;
464 }
465
466 /*
467  * If we have fewer than thresh credits, extend by OCFS2_MAX_TRANS_DATA.
468  * If that fails, restart the transaction & regain write access for the
469  * buffer head which is used for metadata modifications.
470  * Taken from Ext4: extend_or_restart_transaction()
471  */
472 int ocfs2_allocate_extend_trans(handle_t *handle, int thresh)
473 {
474         int status, old_nblks;
475
476         BUG_ON(!handle);
477
478         old_nblks = handle->h_buffer_credits;
479         trace_ocfs2_allocate_extend_trans(old_nblks, thresh);
480
481         if (old_nblks < thresh)
482                 return 0;
483
484         status = jbd2_journal_extend(handle, OCFS2_MAX_TRANS_DATA);
485         if (status < 0) {
486                 mlog_errno(status);
487                 goto bail;
488         }
489
490         if (status > 0) {
491                 status = jbd2_journal_restart(handle, OCFS2_MAX_TRANS_DATA);
492                 if (status < 0)
493                         mlog_errno(status);
494         }
495
496 bail:
497         return status;
498 }
499
500
501 struct ocfs2_triggers {
502         struct jbd2_buffer_trigger_type ot_triggers;
503         int                             ot_offset;
504 };
505
506 static inline struct ocfs2_triggers *to_ocfs2_trigger(struct jbd2_buffer_trigger_type *triggers)
507 {
508         return container_of(triggers, struct ocfs2_triggers, ot_triggers);
509 }
510
511 static void ocfs2_frozen_trigger(struct jbd2_buffer_trigger_type *triggers,
512                                  struct buffer_head *bh,
513                                  void *data, size_t size)
514 {
515         struct ocfs2_triggers *ot = to_ocfs2_trigger(triggers);
516
517         /*
518          * We aren't guaranteed to have the superblock here, so we
519          * must unconditionally compute the ecc data.
520          * __ocfs2_journal_access() will only set the triggers if
521          * metaecc is enabled.
522          */
523         ocfs2_block_check_compute(data, size, data + ot->ot_offset);
524 }
525
526 /*
527  * Quota blocks have their own trigger because the struct ocfs2_block_check
528  * offset depends on the blocksize.
529  */
530 static void ocfs2_dq_frozen_trigger(struct jbd2_buffer_trigger_type *triggers,
531                                  struct buffer_head *bh,
532                                  void *data, size_t size)
533 {
534         struct ocfs2_disk_dqtrailer *dqt =
535                 ocfs2_block_dqtrailer(size, data);
536
537         /*
538          * We aren't guaranteed to have the superblock here, so we
539          * must unconditionally compute the ecc data.
540          * __ocfs2_journal_access() will only set the triggers if
541          * metaecc is enabled.
542          */
543         ocfs2_block_check_compute(data, size, &dqt->dq_check);
544 }
545
546 /*
547  * Directory blocks also have their own trigger because the
548  * struct ocfs2_block_check offset depends on the blocksize.
549  */
550 static void ocfs2_db_frozen_trigger(struct jbd2_buffer_trigger_type *triggers,
551                                  struct buffer_head *bh,
552                                  void *data, size_t size)
553 {
554         struct ocfs2_dir_block_trailer *trailer =
555                 ocfs2_dir_trailer_from_size(size, data);
556
557         /*
558          * We aren't guaranteed to have the superblock here, so we
559          * must unconditionally compute the ecc data.
560          * __ocfs2_journal_access() will only set the triggers if
561          * metaecc is enabled.
562          */
563         ocfs2_block_check_compute(data, size, &trailer->db_check);
564 }
565
566 static void ocfs2_abort_trigger(struct jbd2_buffer_trigger_type *triggers,
567                                 struct buffer_head *bh)
568 {
569         mlog(ML_ERROR,
570              "ocfs2_abort_trigger called by JBD2.  bh = 0x%lx, "
571              "bh->b_blocknr = %llu\n",
572              (unsigned long)bh,
573              (unsigned long long)bh->b_blocknr);
574
575         ocfs2_error(bh->b_bdev->bd_super,
576                     "JBD2 has aborted our journal, ocfs2 cannot continue\n");
577 }
578
579 static struct ocfs2_triggers di_triggers = {
580         .ot_triggers = {
581                 .t_frozen = ocfs2_frozen_trigger,
582                 .t_abort = ocfs2_abort_trigger,
583         },
584         .ot_offset      = offsetof(struct ocfs2_dinode, i_check),
585 };
586
587 static struct ocfs2_triggers eb_triggers = {
588         .ot_triggers = {
589                 .t_frozen = ocfs2_frozen_trigger,
590                 .t_abort = ocfs2_abort_trigger,
591         },
592         .ot_offset      = offsetof(struct ocfs2_extent_block, h_check),
593 };
594
595 static struct ocfs2_triggers rb_triggers = {
596         .ot_triggers = {
597                 .t_frozen = ocfs2_frozen_trigger,
598                 .t_abort = ocfs2_abort_trigger,
599         },
600         .ot_offset      = offsetof(struct ocfs2_refcount_block, rf_check),
601 };
602
603 static struct ocfs2_triggers gd_triggers = {
604         .ot_triggers = {
605                 .t_frozen = ocfs2_frozen_trigger,
606                 .t_abort = ocfs2_abort_trigger,
607         },
608         .ot_offset      = offsetof(struct ocfs2_group_desc, bg_check),
609 };
610
611 static struct ocfs2_triggers db_triggers = {
612         .ot_triggers = {
613                 .t_frozen = ocfs2_db_frozen_trigger,
614                 .t_abort = ocfs2_abort_trigger,
615         },
616 };
617
618 static struct ocfs2_triggers xb_triggers = {
619         .ot_triggers = {
620                 .t_frozen = ocfs2_frozen_trigger,
621                 .t_abort = ocfs2_abort_trigger,
622         },
623         .ot_offset      = offsetof(struct ocfs2_xattr_block, xb_check),
624 };
625
626 static struct ocfs2_triggers dq_triggers = {
627         .ot_triggers = {
628                 .t_frozen = ocfs2_dq_frozen_trigger,
629                 .t_abort = ocfs2_abort_trigger,
630         },
631 };
632
633 static struct ocfs2_triggers dr_triggers = {
634         .ot_triggers = {
635                 .t_frozen = ocfs2_frozen_trigger,
636                 .t_abort = ocfs2_abort_trigger,
637         },
638         .ot_offset      = offsetof(struct ocfs2_dx_root_block, dr_check),
639 };
640
641 static struct ocfs2_triggers dl_triggers = {
642         .ot_triggers = {
643                 .t_frozen = ocfs2_frozen_trigger,
644                 .t_abort = ocfs2_abort_trigger,
645         },
646         .ot_offset      = offsetof(struct ocfs2_dx_leaf, dl_check),
647 };
648
649 static int __ocfs2_journal_access(handle_t *handle,
650                                   struct ocfs2_caching_info *ci,
651                                   struct buffer_head *bh,
652                                   struct ocfs2_triggers *triggers,
653                                   int type)
654 {
655         int status;
656         struct ocfs2_super *osb =
657                 OCFS2_SB(ocfs2_metadata_cache_get_super(ci));
658
659         BUG_ON(!ci || !ci->ci_ops);
660         BUG_ON(!handle);
661         BUG_ON(!bh);
662
663         trace_ocfs2_journal_access(
664                 (unsigned long long)ocfs2_metadata_cache_owner(ci),
665                 (unsigned long long)bh->b_blocknr, type, bh->b_size);
666
667         /* we can safely remove this assertion after testing. */
668         if (!buffer_uptodate(bh)) {
669                 mlog(ML_ERROR, "giving me a buffer that's not uptodate!\n");
670                 mlog(ML_ERROR, "b_blocknr=%llu, b_state=0x%lx\n",
671                      (unsigned long long)bh->b_blocknr, bh->b_state);
672
673                 lock_buffer(bh);
674                 /*
675                  * A previous transaction with a couple of buffer heads fail
676                  * to checkpoint, so all the bhs are marked as BH_Write_EIO.
677                  * For current transaction, the bh is just among those error
678                  * bhs which previous transaction handle. We can't just clear
679                  * its BH_Write_EIO and reuse directly, since other bhs are
680                  * not written to disk yet and that will cause metadata
681                  * inconsistency. So we should set fs read-only to avoid
682                  * further damage.
683                  */
684                 if (buffer_write_io_error(bh) && !buffer_uptodate(bh)) {
685                         unlock_buffer(bh);
686                         return ocfs2_error(osb->sb, "A previous attempt to "
687                                         "write this buffer head failed\n");
688                 }
689                 unlock_buffer(bh);
690         }
691
692         /* Set the current transaction information on the ci so
693          * that the locking code knows whether it can drop it's locks
694          * on this ci or not. We're protected from the commit
695          * thread updating the current transaction id until
696          * ocfs2_commit_trans() because ocfs2_start_trans() took
697          * j_trans_barrier for us. */
698         ocfs2_set_ci_lock_trans(osb->journal, ci);
699
700         ocfs2_metadata_cache_io_lock(ci);
701         switch (type) {
702         case OCFS2_JOURNAL_ACCESS_CREATE:
703         case OCFS2_JOURNAL_ACCESS_WRITE:
704                 status = jbd2_journal_get_write_access(handle, bh);
705                 break;
706
707         case OCFS2_JOURNAL_ACCESS_UNDO:
708                 status = jbd2_journal_get_undo_access(handle, bh);
709                 break;
710
711         default:
712                 status = -EINVAL;
713                 mlog(ML_ERROR, "Unknown access type!\n");
714         }
715         if (!status && ocfs2_meta_ecc(osb) && triggers)
716                 jbd2_journal_set_triggers(bh, &triggers->ot_triggers);
717         ocfs2_metadata_cache_io_unlock(ci);
718
719         if (status < 0)
720                 mlog(ML_ERROR, "Error %d getting %d access to buffer!\n",
721                      status, type);
722
723         return status;
724 }
725
726 int ocfs2_journal_access_di(handle_t *handle, struct ocfs2_caching_info *ci,
727                             struct buffer_head *bh, int type)
728 {
729         return __ocfs2_journal_access(handle, ci, bh, &di_triggers, type);
730 }
731
732 int ocfs2_journal_access_eb(handle_t *handle, struct ocfs2_caching_info *ci,
733                             struct buffer_head *bh, int type)
734 {
735         return __ocfs2_journal_access(handle, ci, bh, &eb_triggers, type);
736 }
737
738 int ocfs2_journal_access_rb(handle_t *handle, struct ocfs2_caching_info *ci,
739                             struct buffer_head *bh, int type)
740 {
741         return __ocfs2_journal_access(handle, ci, bh, &rb_triggers,
742                                       type);
743 }
744
745 int ocfs2_journal_access_gd(handle_t *handle, struct ocfs2_caching_info *ci,
746                             struct buffer_head *bh, int type)
747 {
748         return __ocfs2_journal_access(handle, ci, bh, &gd_triggers, type);
749 }
750
751 int ocfs2_journal_access_db(handle_t *handle, struct ocfs2_caching_info *ci,
752                             struct buffer_head *bh, int type)
753 {
754         return __ocfs2_journal_access(handle, ci, bh, &db_triggers, type);
755 }
756
757 int ocfs2_journal_access_xb(handle_t *handle, struct ocfs2_caching_info *ci,
758                             struct buffer_head *bh, int type)
759 {
760         return __ocfs2_journal_access(handle, ci, bh, &xb_triggers, type);
761 }
762
763 int ocfs2_journal_access_dq(handle_t *handle, struct ocfs2_caching_info *ci,
764                             struct buffer_head *bh, int type)
765 {
766         return __ocfs2_journal_access(handle, ci, bh, &dq_triggers, type);
767 }
768
769 int ocfs2_journal_access_dr(handle_t *handle, struct ocfs2_caching_info *ci,
770                             struct buffer_head *bh, int type)
771 {
772         return __ocfs2_journal_access(handle, ci, bh, &dr_triggers, type);
773 }
774
775 int ocfs2_journal_access_dl(handle_t *handle, struct ocfs2_caching_info *ci,
776                             struct buffer_head *bh, int type)
777 {
778         return __ocfs2_journal_access(handle, ci, bh, &dl_triggers, type);
779 }
780
781 int ocfs2_journal_access(handle_t *handle, struct ocfs2_caching_info *ci,
782                          struct buffer_head *bh, int type)
783 {
784         return __ocfs2_journal_access(handle, ci, bh, NULL, type);
785 }
786
787 void ocfs2_journal_dirty(handle_t *handle, struct buffer_head *bh)
788 {
789         int status;
790
791         trace_ocfs2_journal_dirty((unsigned long long)bh->b_blocknr);
792
793         status = jbd2_journal_dirty_metadata(handle, bh);
794         if (status) {
795                 mlog_errno(status);
796                 if (!is_handle_aborted(handle)) {
797                         journal_t *journal = handle->h_transaction->t_journal;
798                         struct super_block *sb = bh->b_bdev->bd_super;
799
800                         mlog(ML_ERROR, "jbd2_journal_dirty_metadata failed. "
801                                         "Aborting transaction and journal.\n");
802                         handle->h_err = status;
803                         jbd2_journal_abort_handle(handle);
804                         jbd2_journal_abort(journal, status);
805                         ocfs2_abort(sb, "Journal already aborted.\n");
806                 }
807         }
808 }
809
810 #define OCFS2_DEFAULT_COMMIT_INTERVAL   (HZ * JBD2_DEFAULT_MAX_COMMIT_AGE)
811
812 void ocfs2_set_journal_params(struct ocfs2_super *osb)
813 {
814         journal_t *journal = osb->journal->j_journal;
815         unsigned long commit_interval = OCFS2_DEFAULT_COMMIT_INTERVAL;
816
817         if (osb->osb_commit_interval)
818                 commit_interval = osb->osb_commit_interval;
819
820         write_lock(&journal->j_state_lock);
821         journal->j_commit_interval = commit_interval;
822         if (osb->s_mount_opt & OCFS2_MOUNT_BARRIER)
823                 journal->j_flags |= JBD2_BARRIER;
824         else
825                 journal->j_flags &= ~JBD2_BARRIER;
826         write_unlock(&journal->j_state_lock);
827 }
828
829 int ocfs2_journal_init(struct ocfs2_journal *journal, int *dirty)
830 {
831         int status = -1;
832         struct inode *inode = NULL; /* the journal inode */
833         journal_t *j_journal = NULL;
834         struct ocfs2_dinode *di = NULL;
835         struct buffer_head *bh = NULL;
836         struct ocfs2_super *osb;
837         int inode_lock = 0;
838
839         BUG_ON(!journal);
840
841         osb = journal->j_osb;
842
843         /* already have the inode for our journal */
844         inode = ocfs2_get_system_file_inode(osb, JOURNAL_SYSTEM_INODE,
845                                             osb->slot_num);
846         if (inode == NULL) {
847                 status = -EACCES;
848                 mlog_errno(status);
849                 goto done;
850         }
851         if (is_bad_inode(inode)) {
852                 mlog(ML_ERROR, "access error (bad inode)\n");
853                 iput(inode);
854                 inode = NULL;
855                 status = -EACCES;
856                 goto done;
857         }
858
859         SET_INODE_JOURNAL(inode);
860         OCFS2_I(inode)->ip_open_count++;
861
862         /* Skip recovery waits here - journal inode metadata never
863          * changes in a live cluster so it can be considered an
864          * exception to the rule. */
865         status = ocfs2_inode_lock_full(inode, &bh, 1, OCFS2_META_LOCK_RECOVERY);
866         if (status < 0) {
867                 if (status != -ERESTARTSYS)
868                         mlog(ML_ERROR, "Could not get lock on journal!\n");
869                 goto done;
870         }
871
872         inode_lock = 1;
873         di = (struct ocfs2_dinode *)bh->b_data;
874
875         if (i_size_read(inode) <  OCFS2_MIN_JOURNAL_SIZE) {
876                 mlog(ML_ERROR, "Journal file size (%lld) is too small!\n",
877                      i_size_read(inode));
878                 status = -EINVAL;
879                 goto done;
880         }
881
882         trace_ocfs2_journal_init(i_size_read(inode),
883                                  (unsigned long long)inode->i_blocks,
884                                  OCFS2_I(inode)->ip_clusters);
885
886         /* call the kernels journal init function now */
887         j_journal = jbd2_journal_init_inode(inode);
888         if (j_journal == NULL) {
889                 mlog(ML_ERROR, "Linux journal layer error\n");
890                 status = -EINVAL;
891                 goto done;
892         }
893
894         trace_ocfs2_journal_init_maxlen(j_journal->j_maxlen);
895
896         *dirty = (le32_to_cpu(di->id1.journal1.ij_flags) &
897                   OCFS2_JOURNAL_DIRTY_FL);
898
899         journal->j_journal = j_journal;
900         journal->j_inode = inode;
901         journal->j_bh = bh;
902
903         ocfs2_set_journal_params(osb);
904
905         journal->j_state = OCFS2_JOURNAL_LOADED;
906
907         status = 0;
908 done:
909         if (status < 0) {
910                 if (inode_lock)
911                         ocfs2_inode_unlock(inode, 1);
912                 brelse(bh);
913                 if (inode) {
914                         OCFS2_I(inode)->ip_open_count--;
915                         iput(inode);
916                 }
917         }
918
919         return status;
920 }
921
922 static void ocfs2_bump_recovery_generation(struct ocfs2_dinode *di)
923 {
924         le32_add_cpu(&(di->id1.journal1.ij_recovery_generation), 1);
925 }
926
927 static u32 ocfs2_get_recovery_generation(struct ocfs2_dinode *di)
928 {
929         return le32_to_cpu(di->id1.journal1.ij_recovery_generation);
930 }
931
932 static int ocfs2_journal_toggle_dirty(struct ocfs2_super *osb,
933                                       int dirty, int replayed)
934 {
935         int status;
936         unsigned int flags;
937         struct ocfs2_journal *journal = osb->journal;
938         struct buffer_head *bh = journal->j_bh;
939         struct ocfs2_dinode *fe;
940
941         fe = (struct ocfs2_dinode *)bh->b_data;
942
943         /* The journal bh on the osb always comes from ocfs2_journal_init()
944          * and was validated there inside ocfs2_inode_lock_full().  It's a
945          * code bug if we mess it up. */
946         BUG_ON(!OCFS2_IS_VALID_DINODE(fe));
947
948         flags = le32_to_cpu(fe->id1.journal1.ij_flags);
949         if (dirty)
950                 flags |= OCFS2_JOURNAL_DIRTY_FL;
951         else
952                 flags &= ~OCFS2_JOURNAL_DIRTY_FL;
953         fe->id1.journal1.ij_flags = cpu_to_le32(flags);
954
955         if (replayed)
956                 ocfs2_bump_recovery_generation(fe);
957
958         ocfs2_compute_meta_ecc(osb->sb, bh->b_data, &fe->i_check);
959         status = ocfs2_write_block(osb, bh, INODE_CACHE(journal->j_inode));
960         if (status < 0)
961                 mlog_errno(status);
962
963         return status;
964 }
965
966 /*
967  * If the journal has been kmalloc'd it needs to be freed after this
968  * call.
969  */
970 void ocfs2_journal_shutdown(struct ocfs2_super *osb)
971 {
972         struct ocfs2_journal *journal = NULL;
973         int status = 0;
974         struct inode *inode = NULL;
975         int num_running_trans = 0;
976
977         BUG_ON(!osb);
978
979         journal = osb->journal;
980         if (!journal)
981                 goto done;
982
983         inode = journal->j_inode;
984
985         if (journal->j_state != OCFS2_JOURNAL_LOADED)
986                 goto done;
987
988         /* need to inc inode use count - jbd2_journal_destroy will iput. */
989         if (!igrab(inode))
990                 BUG();
991
992         num_running_trans = atomic_read(&(osb->journal->j_num_trans));
993         trace_ocfs2_journal_shutdown(num_running_trans);
994
995         /* Do a commit_cache here. It will flush our journal, *and*
996          * release any locks that are still held.
997          * set the SHUTDOWN flag and release the trans lock.
998          * the commit thread will take the trans lock for us below. */
999         journal->j_state = OCFS2_JOURNAL_IN_SHUTDOWN;
1000
1001         /* The OCFS2_JOURNAL_IN_SHUTDOWN will signal to commit_cache to not
1002          * drop the trans_lock (which we want to hold until we
1003          * completely destroy the journal. */
1004         if (osb->commit_task) {
1005                 /* Wait for the commit thread */
1006                 trace_ocfs2_journal_shutdown_wait(osb->commit_task);
1007                 kthread_stop(osb->commit_task);
1008                 osb->commit_task = NULL;
1009         }
1010
1011         BUG_ON(atomic_read(&(osb->journal->j_num_trans)) != 0);
1012
1013         if (ocfs2_mount_local(osb)) {
1014                 jbd2_journal_lock_updates(journal->j_journal);
1015                 status = jbd2_journal_flush(journal->j_journal);
1016                 jbd2_journal_unlock_updates(journal->j_journal);
1017                 if (status < 0)
1018                         mlog_errno(status);
1019         }
1020
1021         /* Shutdown the kernel journal system */
1022         if (!jbd2_journal_destroy(journal->j_journal) && !status) {
1023                 /*
1024                  * Do not toggle if flush was unsuccessful otherwise
1025                  * will leave dirty metadata in a "clean" journal
1026                  */
1027                 status = ocfs2_journal_toggle_dirty(osb, 0, 0);
1028                 if (status < 0)
1029                         mlog_errno(status);
1030         }
1031         journal->j_journal = NULL;
1032
1033         OCFS2_I(inode)->ip_open_count--;
1034
1035         /* unlock our journal */
1036         ocfs2_inode_unlock(inode, 1);
1037
1038         brelse(journal->j_bh);
1039         journal->j_bh = NULL;
1040
1041         journal->j_state = OCFS2_JOURNAL_FREE;
1042
1043 //      up_write(&journal->j_trans_barrier);
1044 done:
1045         iput(inode);
1046 }
1047
1048 static void ocfs2_clear_journal_error(struct super_block *sb,
1049                                       journal_t *journal,
1050                                       int slot)
1051 {
1052         int olderr;
1053
1054         olderr = jbd2_journal_errno(journal);
1055         if (olderr) {
1056                 mlog(ML_ERROR, "File system error %d recorded in "
1057                      "journal %u.\n", olderr, slot);
1058                 mlog(ML_ERROR, "File system on device %s needs checking.\n",
1059                      sb->s_id);
1060
1061                 jbd2_journal_ack_err(journal);
1062                 jbd2_journal_clear_err(journal);
1063         }
1064 }
1065
1066 int ocfs2_journal_load(struct ocfs2_journal *journal, int local, int replayed)
1067 {
1068         int status = 0;
1069         struct ocfs2_super *osb;
1070
1071         BUG_ON(!journal);
1072
1073         osb = journal->j_osb;
1074
1075         status = jbd2_journal_load(journal->j_journal);
1076         if (status < 0) {
1077                 mlog(ML_ERROR, "Failed to load journal!\n");
1078                 goto done;
1079         }
1080
1081         ocfs2_clear_journal_error(osb->sb, journal->j_journal, osb->slot_num);
1082
1083         if (replayed) {
1084                 jbd2_journal_lock_updates(journal->j_journal);
1085                 status = jbd2_journal_flush(journal->j_journal);
1086                 jbd2_journal_unlock_updates(journal->j_journal);
1087                 if (status < 0)
1088                         mlog_errno(status);
1089         }
1090
1091         status = ocfs2_journal_toggle_dirty(osb, 1, replayed);
1092         if (status < 0) {
1093                 mlog_errno(status);
1094                 goto done;
1095         }
1096
1097         /* Launch the commit thread */
1098         if (!local) {
1099                 osb->commit_task = kthread_run(ocfs2_commit_thread, osb,
1100                                 "ocfs2cmt-%s", osb->uuid_str);
1101                 if (IS_ERR(osb->commit_task)) {
1102                         status = PTR_ERR(osb->commit_task);
1103                         osb->commit_task = NULL;
1104                         mlog(ML_ERROR, "unable to launch ocfs2commit thread, "
1105                              "error=%d", status);
1106                         goto done;
1107                 }
1108         } else
1109                 osb->commit_task = NULL;
1110
1111 done:
1112         return status;
1113 }
1114
1115
1116 /* 'full' flag tells us whether we clear out all blocks or if we just
1117  * mark the journal clean */
1118 int ocfs2_journal_wipe(struct ocfs2_journal *journal, int full)
1119 {
1120         int status;
1121
1122         BUG_ON(!journal);
1123
1124         status = jbd2_journal_wipe(journal->j_journal, full);
1125         if (status < 0) {
1126                 mlog_errno(status);
1127                 goto bail;
1128         }
1129
1130         status = ocfs2_journal_toggle_dirty(journal->j_osb, 0, 0);
1131         if (status < 0)
1132                 mlog_errno(status);
1133
1134 bail:
1135         return status;
1136 }
1137
1138 static int ocfs2_recovery_completed(struct ocfs2_super *osb)
1139 {
1140         int empty;
1141         struct ocfs2_recovery_map *rm = osb->recovery_map;
1142
1143         spin_lock(&osb->osb_lock);
1144         empty = (rm->rm_used == 0);
1145         spin_unlock(&osb->osb_lock);
1146
1147         return empty;
1148 }
1149
1150 void ocfs2_wait_for_recovery(struct ocfs2_super *osb)
1151 {
1152         wait_event(osb->recovery_event, ocfs2_recovery_completed(osb));
1153 }
1154
1155 /*
1156  * JBD Might read a cached version of another nodes journal file. We
1157  * don't want this as this file changes often and we get no
1158  * notification on those changes. The only way to be sure that we've
1159  * got the most up to date version of those blocks then is to force
1160  * read them off disk. Just searching through the buffer cache won't
1161  * work as there may be pages backing this file which are still marked
1162  * up to date. We know things can't change on this file underneath us
1163  * as we have the lock by now :)
1164  */
1165 static int ocfs2_force_read_journal(struct inode *inode)
1166 {
1167         int status = 0;
1168         int i;
1169         u64 v_blkno, p_blkno, p_blocks, num_blocks;
1170         struct buffer_head *bh = NULL;
1171         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1172
1173         num_blocks = ocfs2_blocks_for_bytes(inode->i_sb, i_size_read(inode));
1174         v_blkno = 0;
1175         while (v_blkno < num_blocks) {
1176                 status = ocfs2_extent_map_get_blocks(inode, v_blkno,
1177                                                      &p_blkno, &p_blocks, NULL);
1178                 if (status < 0) {
1179                         mlog_errno(status);
1180                         goto bail;
1181                 }
1182
1183                 for (i = 0; i < p_blocks; i++, p_blkno++) {
1184                         bh = __find_get_block(osb->sb->s_bdev, p_blkno,
1185                                         osb->sb->s_blocksize);
1186                         /* block not cached. */
1187                         if (!bh)
1188                                 continue;
1189
1190                         brelse(bh);
1191                         bh = NULL;
1192                         /* We are reading journal data which should not
1193                          * be put in the uptodate cache.
1194                          */
1195                         status = ocfs2_read_blocks_sync(osb, p_blkno, 1, &bh);
1196                         if (status < 0) {
1197                                 mlog_errno(status);
1198                                 goto bail;
1199                         }
1200
1201                         brelse(bh);
1202                         bh = NULL;
1203                 }
1204
1205                 v_blkno += p_blocks;
1206         }
1207
1208 bail:
1209         return status;
1210 }
1211
1212 struct ocfs2_la_recovery_item {
1213         struct list_head        lri_list;
1214         int                     lri_slot;
1215         struct ocfs2_dinode     *lri_la_dinode;
1216         struct ocfs2_dinode     *lri_tl_dinode;
1217         struct ocfs2_quota_recovery *lri_qrec;
1218         enum ocfs2_orphan_reco_type  lri_orphan_reco_type;
1219 };
1220
1221 /* Does the second half of the recovery process. By this point, the
1222  * node is marked clean and can actually be considered recovered,
1223  * hence it's no longer in the recovery map, but there's still some
1224  * cleanup we can do which shouldn't happen within the recovery thread
1225  * as locking in that context becomes very difficult if we are to take
1226  * recovering nodes into account.
1227  *
1228  * NOTE: This function can and will sleep on recovery of other nodes
1229  * during cluster locking, just like any other ocfs2 process.
1230  */
1231 void ocfs2_complete_recovery(struct work_struct *work)
1232 {
1233         int ret = 0;
1234         struct ocfs2_journal *journal =
1235                 container_of(work, struct ocfs2_journal, j_recovery_work);
1236         struct ocfs2_super *osb = journal->j_osb;
1237         struct ocfs2_dinode *la_dinode, *tl_dinode;
1238         struct ocfs2_la_recovery_item *item, *n;
1239         struct ocfs2_quota_recovery *qrec;
1240         enum ocfs2_orphan_reco_type orphan_reco_type;
1241         LIST_HEAD(tmp_la_list);
1242
1243         trace_ocfs2_complete_recovery(
1244                 (unsigned long long)OCFS2_I(journal->j_inode)->ip_blkno);
1245
1246         spin_lock(&journal->j_lock);
1247         list_splice_init(&journal->j_la_cleanups, &tmp_la_list);
1248         spin_unlock(&journal->j_lock);
1249
1250         list_for_each_entry_safe(item, n, &tmp_la_list, lri_list) {
1251                 list_del_init(&item->lri_list);
1252
1253                 ocfs2_wait_on_quotas(osb);
1254
1255                 la_dinode = item->lri_la_dinode;
1256                 tl_dinode = item->lri_tl_dinode;
1257                 qrec = item->lri_qrec;
1258                 orphan_reco_type = item->lri_orphan_reco_type;
1259
1260                 trace_ocfs2_complete_recovery_slot(item->lri_slot,
1261                         la_dinode ? le64_to_cpu(la_dinode->i_blkno) : 0,
1262                         tl_dinode ? le64_to_cpu(tl_dinode->i_blkno) : 0,
1263                         qrec);
1264
1265                 if (la_dinode) {
1266                         ret = ocfs2_complete_local_alloc_recovery(osb,
1267                                                                   la_dinode);
1268                         if (ret < 0)
1269                                 mlog_errno(ret);
1270
1271                         kfree(la_dinode);
1272                 }
1273
1274                 if (tl_dinode) {
1275                         ret = ocfs2_complete_truncate_log_recovery(osb,
1276                                                                    tl_dinode);
1277                         if (ret < 0)
1278                                 mlog_errno(ret);
1279
1280                         kfree(tl_dinode);
1281                 }
1282
1283                 ret = ocfs2_recover_orphans(osb, item->lri_slot,
1284                                 orphan_reco_type);
1285                 if (ret < 0)
1286                         mlog_errno(ret);
1287
1288                 if (qrec) {
1289                         ret = ocfs2_finish_quota_recovery(osb, qrec,
1290                                                           item->lri_slot);
1291                         if (ret < 0)
1292                                 mlog_errno(ret);
1293                         /* Recovery info is already freed now */
1294                 }
1295
1296                 kfree(item);
1297         }
1298
1299         trace_ocfs2_complete_recovery_end(ret);
1300 }
1301
1302 /* NOTE: This function always eats your references to la_dinode and
1303  * tl_dinode, either manually on error, or by passing them to
1304  * ocfs2_complete_recovery */
1305 static void ocfs2_queue_recovery_completion(struct ocfs2_journal *journal,
1306                                             int slot_num,
1307                                             struct ocfs2_dinode *la_dinode,
1308                                             struct ocfs2_dinode *tl_dinode,
1309                                             struct ocfs2_quota_recovery *qrec,
1310                                             enum ocfs2_orphan_reco_type orphan_reco_type)
1311 {
1312         struct ocfs2_la_recovery_item *item;
1313
1314         item = kmalloc(sizeof(struct ocfs2_la_recovery_item), GFP_NOFS);
1315         if (!item) {
1316                 /* Though we wish to avoid it, we are in fact safe in
1317                  * skipping local alloc cleanup as fsck.ocfs2 is more
1318                  * than capable of reclaiming unused space. */
1319                 kfree(la_dinode);
1320                 kfree(tl_dinode);
1321
1322                 if (qrec)
1323                         ocfs2_free_quota_recovery(qrec);
1324
1325                 mlog_errno(-ENOMEM);
1326                 return;
1327         }
1328
1329         INIT_LIST_HEAD(&item->lri_list);
1330         item->lri_la_dinode = la_dinode;
1331         item->lri_slot = slot_num;
1332         item->lri_tl_dinode = tl_dinode;
1333         item->lri_qrec = qrec;
1334         item->lri_orphan_reco_type = orphan_reco_type;
1335
1336         spin_lock(&journal->j_lock);
1337         list_add_tail(&item->lri_list, &journal->j_la_cleanups);
1338         queue_work(journal->j_osb->ocfs2_wq, &journal->j_recovery_work);
1339         spin_unlock(&journal->j_lock);
1340 }
1341
1342 /* Called by the mount code to queue recovery the last part of
1343  * recovery for it's own and offline slot(s). */
1344 void ocfs2_complete_mount_recovery(struct ocfs2_super *osb)
1345 {
1346         struct ocfs2_journal *journal = osb->journal;
1347
1348         if (ocfs2_is_hard_readonly(osb))
1349                 return;
1350
1351         /* No need to queue up our truncate_log as regular cleanup will catch
1352          * that */
1353         ocfs2_queue_recovery_completion(journal, osb->slot_num,
1354                                         osb->local_alloc_copy, NULL, NULL,
1355                                         ORPHAN_NEED_TRUNCATE);
1356         ocfs2_schedule_truncate_log_flush(osb, 0);
1357
1358         osb->local_alloc_copy = NULL;
1359         osb->dirty = 0;
1360
1361         /* queue to recover orphan slots for all offline slots */
1362         ocfs2_replay_map_set_state(osb, REPLAY_NEEDED);
1363         ocfs2_queue_replay_slots(osb, ORPHAN_NEED_TRUNCATE);
1364         ocfs2_free_replay_slots(osb);
1365 }
1366
1367 void ocfs2_complete_quota_recovery(struct ocfs2_super *osb)
1368 {
1369         if (osb->quota_rec) {
1370                 ocfs2_queue_recovery_completion(osb->journal,
1371                                                 osb->slot_num,
1372                                                 NULL,
1373                                                 NULL,
1374                                                 osb->quota_rec,
1375                                                 ORPHAN_NEED_TRUNCATE);
1376                 osb->quota_rec = NULL;
1377         }
1378 }
1379
1380 static int __ocfs2_recovery_thread(void *arg)
1381 {
1382         int status, node_num, slot_num;
1383         struct ocfs2_super *osb = arg;
1384         struct ocfs2_recovery_map *rm = osb->recovery_map;
1385         int *rm_quota = NULL;
1386         int rm_quota_used = 0, i;
1387         struct ocfs2_quota_recovery *qrec;
1388
1389         status = ocfs2_wait_on_mount(osb);
1390         if (status < 0) {
1391                 goto bail;
1392         }
1393
1394         rm_quota = kzalloc(osb->max_slots * sizeof(int), GFP_NOFS);
1395         if (!rm_quota) {
1396                 status = -ENOMEM;
1397                 goto bail;
1398         }
1399 restart:
1400         status = ocfs2_super_lock(osb, 1);
1401         if (status < 0) {
1402                 mlog_errno(status);
1403                 goto bail;
1404         }
1405
1406         status = ocfs2_compute_replay_slots(osb);
1407         if (status < 0)
1408                 mlog_errno(status);
1409
1410         /* queue recovery for our own slot */
1411         ocfs2_queue_recovery_completion(osb->journal, osb->slot_num, NULL,
1412                                         NULL, NULL, ORPHAN_NO_NEED_TRUNCATE);
1413
1414         spin_lock(&osb->osb_lock);
1415         while (rm->rm_used) {
1416                 /* It's always safe to remove entry zero, as we won't
1417                  * clear it until ocfs2_recover_node() has succeeded. */
1418                 node_num = rm->rm_entries[0];
1419                 spin_unlock(&osb->osb_lock);
1420                 slot_num = ocfs2_node_num_to_slot(osb, node_num);
1421                 trace_ocfs2_recovery_thread_node(node_num, slot_num);
1422                 if (slot_num == -ENOENT) {
1423                         status = 0;
1424                         goto skip_recovery;
1425                 }
1426
1427                 /* It is a bit subtle with quota recovery. We cannot do it
1428                  * immediately because we have to obtain cluster locks from
1429                  * quota files and we also don't want to just skip it because
1430                  * then quota usage would be out of sync until some node takes
1431                  * the slot. So we remember which nodes need quota recovery
1432                  * and when everything else is done, we recover quotas. */
1433                 for (i = 0; i < rm_quota_used && rm_quota[i] != slot_num; i++);
1434                 if (i == rm_quota_used)
1435                         rm_quota[rm_quota_used++] = slot_num;
1436
1437                 status = ocfs2_recover_node(osb, node_num, slot_num);
1438 skip_recovery:
1439                 if (!status) {
1440                         ocfs2_recovery_map_clear(osb, node_num);
1441                 } else {
1442                         mlog(ML_ERROR,
1443                              "Error %d recovering node %d on device (%u,%u)!\n",
1444                              status, node_num,
1445                              MAJOR(osb->sb->s_dev), MINOR(osb->sb->s_dev));
1446                         mlog(ML_ERROR, "Volume requires unmount.\n");
1447                 }
1448
1449                 spin_lock(&osb->osb_lock);
1450         }
1451         spin_unlock(&osb->osb_lock);
1452         trace_ocfs2_recovery_thread_end(status);
1453
1454         /* Refresh all journal recovery generations from disk */
1455         status = ocfs2_check_journals_nolocks(osb);
1456         status = (status == -EROFS) ? 0 : status;
1457         if (status < 0)
1458                 mlog_errno(status);
1459
1460         /* Now it is right time to recover quotas... We have to do this under
1461          * superblock lock so that no one can start using the slot (and crash)
1462          * before we recover it */
1463         for (i = 0; i < rm_quota_used; i++) {
1464                 qrec = ocfs2_begin_quota_recovery(osb, rm_quota[i]);
1465                 if (IS_ERR(qrec)) {
1466                         status = PTR_ERR(qrec);
1467                         mlog_errno(status);
1468                         continue;
1469                 }
1470                 ocfs2_queue_recovery_completion(osb->journal, rm_quota[i],
1471                                                 NULL, NULL, qrec,
1472                                                 ORPHAN_NEED_TRUNCATE);
1473         }
1474
1475         ocfs2_super_unlock(osb, 1);
1476
1477         /* queue recovery for offline slots */
1478         ocfs2_queue_replay_slots(osb, ORPHAN_NEED_TRUNCATE);
1479
1480 bail:
1481         mutex_lock(&osb->recovery_lock);
1482         if (!status && !ocfs2_recovery_completed(osb)) {
1483                 mutex_unlock(&osb->recovery_lock);
1484                 goto restart;
1485         }
1486
1487         ocfs2_free_replay_slots(osb);
1488         osb->recovery_thread_task = NULL;
1489         mb(); /* sync with ocfs2_recovery_thread_running */
1490         wake_up(&osb->recovery_event);
1491
1492         mutex_unlock(&osb->recovery_lock);
1493
1494         kfree(rm_quota);
1495
1496         /* no one is callint kthread_stop() for us so the kthread() api
1497          * requires that we call do_exit().  And it isn't exported, but
1498          * complete_and_exit() seems to be a minimal wrapper around it. */
1499         complete_and_exit(NULL, status);
1500 }
1501
1502 void ocfs2_recovery_thread(struct ocfs2_super *osb, int node_num)
1503 {
1504         mutex_lock(&osb->recovery_lock);
1505
1506         trace_ocfs2_recovery_thread(node_num, osb->node_num,
1507                 osb->disable_recovery, osb->recovery_thread_task,
1508                 osb->disable_recovery ?
1509                 -1 : ocfs2_recovery_map_set(osb, node_num));
1510
1511         if (osb->disable_recovery)
1512                 goto out;
1513
1514         if (osb->recovery_thread_task)
1515                 goto out;
1516
1517         osb->recovery_thread_task =  kthread_run(__ocfs2_recovery_thread, osb,
1518                         "ocfs2rec-%s", osb->uuid_str);
1519         if (IS_ERR(osb->recovery_thread_task)) {
1520                 mlog_errno((int)PTR_ERR(osb->recovery_thread_task));
1521                 osb->recovery_thread_task = NULL;
1522         }
1523
1524 out:
1525         mutex_unlock(&osb->recovery_lock);
1526         wake_up(&osb->recovery_event);
1527 }
1528
1529 static int ocfs2_read_journal_inode(struct ocfs2_super *osb,
1530                                     int slot_num,
1531                                     struct buffer_head **bh,
1532                                     struct inode **ret_inode)
1533 {
1534         int status = -EACCES;
1535         struct inode *inode = NULL;
1536
1537         BUG_ON(slot_num >= osb->max_slots);
1538
1539         inode = ocfs2_get_system_file_inode(osb, JOURNAL_SYSTEM_INODE,
1540                                             slot_num);
1541         if (!inode || is_bad_inode(inode)) {
1542                 mlog_errno(status);
1543                 goto bail;
1544         }
1545         SET_INODE_JOURNAL(inode);
1546
1547         status = ocfs2_read_inode_block_full(inode, bh, OCFS2_BH_IGNORE_CACHE);
1548         if (status < 0) {
1549                 mlog_errno(status);
1550                 goto bail;
1551         }
1552
1553         status = 0;
1554
1555 bail:
1556         if (inode) {
1557                 if (status || !ret_inode)
1558                         iput(inode);
1559                 else
1560                         *ret_inode = inode;
1561         }
1562         return status;
1563 }
1564
1565 /* Does the actual journal replay and marks the journal inode as
1566  * clean. Will only replay if the journal inode is marked dirty. */
1567 static int ocfs2_replay_journal(struct ocfs2_super *osb,
1568                                 int node_num,
1569                                 int slot_num)
1570 {
1571         int status;
1572         int got_lock = 0;
1573         unsigned int flags;
1574         struct inode *inode = NULL;
1575         struct ocfs2_dinode *fe;
1576         journal_t *journal = NULL;
1577         struct buffer_head *bh = NULL;
1578         u32 slot_reco_gen;
1579
1580         status = ocfs2_read_journal_inode(osb, slot_num, &bh, &inode);
1581         if (status) {
1582                 mlog_errno(status);
1583                 goto done;
1584         }
1585
1586         fe = (struct ocfs2_dinode *)bh->b_data;
1587         slot_reco_gen = ocfs2_get_recovery_generation(fe);
1588         brelse(bh);
1589         bh = NULL;
1590
1591         /*
1592          * As the fs recovery is asynchronous, there is a small chance that
1593          * another node mounted (and recovered) the slot before the recovery
1594          * thread could get the lock. To handle that, we dirty read the journal
1595          * inode for that slot to get the recovery generation. If it is
1596          * different than what we expected, the slot has been recovered.
1597          * If not, it needs recovery.
1598          */
1599         if (osb->slot_recovery_generations[slot_num] != slot_reco_gen) {
1600                 trace_ocfs2_replay_journal_recovered(slot_num,
1601                      osb->slot_recovery_generations[slot_num], slot_reco_gen);
1602                 osb->slot_recovery_generations[slot_num] = slot_reco_gen;
1603                 status = -EBUSY;
1604                 goto done;
1605         }
1606
1607         /* Continue with recovery as the journal has not yet been recovered */
1608
1609         status = ocfs2_inode_lock_full(inode, &bh, 1, OCFS2_META_LOCK_RECOVERY);
1610         if (status < 0) {
1611                 trace_ocfs2_replay_journal_lock_err(status);
1612                 if (status != -ERESTARTSYS)
1613                         mlog(ML_ERROR, "Could not lock journal!\n");
1614                 goto done;
1615         }
1616         got_lock = 1;
1617
1618         fe = (struct ocfs2_dinode *) bh->b_data;
1619
1620         flags = le32_to_cpu(fe->id1.journal1.ij_flags);
1621         slot_reco_gen = ocfs2_get_recovery_generation(fe);
1622
1623         if (!(flags & OCFS2_JOURNAL_DIRTY_FL)) {
1624                 trace_ocfs2_replay_journal_skip(node_num);
1625                 /* Refresh recovery generation for the slot */
1626                 osb->slot_recovery_generations[slot_num] = slot_reco_gen;
1627                 goto done;
1628         }
1629
1630         /* we need to run complete recovery for offline orphan slots */
1631         ocfs2_replay_map_set_state(osb, REPLAY_NEEDED);
1632
1633         printk(KERN_NOTICE "ocfs2: Begin replay journal (node %d, slot %d) on "\
1634                "device (%u,%u)\n", node_num, slot_num, MAJOR(osb->sb->s_dev),
1635                MINOR(osb->sb->s_dev));
1636
1637         OCFS2_I(inode)->ip_clusters = le32_to_cpu(fe->i_clusters);
1638
1639         status = ocfs2_force_read_journal(inode);
1640         if (status < 0) {
1641                 mlog_errno(status);
1642                 goto done;
1643         }
1644
1645         journal = jbd2_journal_init_inode(inode);
1646         if (journal == NULL) {
1647                 mlog(ML_ERROR, "Linux journal layer error\n");
1648                 status = -EIO;
1649                 goto done;
1650         }
1651
1652         status = jbd2_journal_load(journal);
1653         if (status < 0) {
1654                 mlog_errno(status);
1655                 if (!igrab(inode))
1656                         BUG();
1657                 jbd2_journal_destroy(journal);
1658                 goto done;
1659         }
1660
1661         ocfs2_clear_journal_error(osb->sb, journal, slot_num);
1662
1663         /* wipe the journal */
1664         jbd2_journal_lock_updates(journal);
1665         status = jbd2_journal_flush(journal);
1666         jbd2_journal_unlock_updates(journal);
1667         if (status < 0)
1668                 mlog_errno(status);
1669
1670         /* This will mark the node clean */
1671         flags = le32_to_cpu(fe->id1.journal1.ij_flags);
1672         flags &= ~OCFS2_JOURNAL_DIRTY_FL;
1673         fe->id1.journal1.ij_flags = cpu_to_le32(flags);
1674
1675         /* Increment recovery generation to indicate successful recovery */
1676         ocfs2_bump_recovery_generation(fe);
1677         osb->slot_recovery_generations[slot_num] =
1678                                         ocfs2_get_recovery_generation(fe);
1679
1680         ocfs2_compute_meta_ecc(osb->sb, bh->b_data, &fe->i_check);
1681         status = ocfs2_write_block(osb, bh, INODE_CACHE(inode));
1682         if (status < 0)
1683                 mlog_errno(status);
1684
1685         if (!igrab(inode))
1686                 BUG();
1687
1688         jbd2_journal_destroy(journal);
1689
1690         printk(KERN_NOTICE "ocfs2: End replay journal (node %d, slot %d) on "\
1691                "device (%u,%u)\n", node_num, slot_num, MAJOR(osb->sb->s_dev),
1692                MINOR(osb->sb->s_dev));
1693 done:
1694         /* drop the lock on this nodes journal */
1695         if (got_lock)
1696                 ocfs2_inode_unlock(inode, 1);
1697
1698         iput(inode);
1699         brelse(bh);
1700
1701         return status;
1702 }
1703
1704 /*
1705  * Do the most important parts of node recovery:
1706  *  - Replay it's journal
1707  *  - Stamp a clean local allocator file
1708  *  - Stamp a clean truncate log
1709  *  - Mark the node clean
1710  *
1711  * If this function completes without error, a node in OCFS2 can be
1712  * said to have been safely recovered. As a result, failure during the
1713  * second part of a nodes recovery process (local alloc recovery) is
1714  * far less concerning.
1715  */
1716 static int ocfs2_recover_node(struct ocfs2_super *osb,
1717                               int node_num, int slot_num)
1718 {
1719         int status = 0;
1720         struct ocfs2_dinode *la_copy = NULL;
1721         struct ocfs2_dinode *tl_copy = NULL;
1722
1723         trace_ocfs2_recover_node(node_num, slot_num, osb->node_num);
1724
1725         /* Should not ever be called to recover ourselves -- in that
1726          * case we should've called ocfs2_journal_load instead. */
1727         BUG_ON(osb->node_num == node_num);
1728
1729         status = ocfs2_replay_journal(osb, node_num, slot_num);
1730         if (status < 0) {
1731                 if (status == -EBUSY) {
1732                         trace_ocfs2_recover_node_skip(slot_num, node_num);
1733                         status = 0;
1734                         goto done;
1735                 }
1736                 mlog_errno(status);
1737                 goto done;
1738         }
1739
1740         /* Stamp a clean local alloc file AFTER recovering the journal... */
1741         status = ocfs2_begin_local_alloc_recovery(osb, slot_num, &la_copy);
1742         if (status < 0) {
1743                 mlog_errno(status);
1744                 goto done;
1745         }
1746
1747         /* An error from begin_truncate_log_recovery is not
1748          * serious enough to warrant halting the rest of
1749          * recovery. */
1750         status = ocfs2_begin_truncate_log_recovery(osb, slot_num, &tl_copy);
1751         if (status < 0)
1752                 mlog_errno(status);
1753
1754         /* Likewise, this would be a strange but ultimately not so
1755          * harmful place to get an error... */
1756         status = ocfs2_clear_slot(osb, slot_num);
1757         if (status < 0)
1758                 mlog_errno(status);
1759
1760         /* This will kfree the memory pointed to by la_copy and tl_copy */
1761         ocfs2_queue_recovery_completion(osb->journal, slot_num, la_copy,
1762                                         tl_copy, NULL, ORPHAN_NEED_TRUNCATE);
1763
1764         status = 0;
1765 done:
1766
1767         return status;
1768 }
1769
1770 /* Test node liveness by trylocking his journal. If we get the lock,
1771  * we drop it here. Return 0 if we got the lock, -EAGAIN if node is
1772  * still alive (we couldn't get the lock) and < 0 on error. */
1773 static int ocfs2_trylock_journal(struct ocfs2_super *osb,
1774                                  int slot_num)
1775 {
1776         int status, flags;
1777         struct inode *inode = NULL;
1778
1779         inode = ocfs2_get_system_file_inode(osb, JOURNAL_SYSTEM_INODE,
1780                                             slot_num);
1781         if (inode == NULL) {
1782                 mlog(ML_ERROR, "access error\n");
1783                 status = -EACCES;
1784                 goto bail;
1785         }
1786         if (is_bad_inode(inode)) {
1787                 mlog(ML_ERROR, "access error (bad inode)\n");
1788                 iput(inode);
1789                 inode = NULL;
1790                 status = -EACCES;
1791                 goto bail;
1792         }
1793         SET_INODE_JOURNAL(inode);
1794
1795         flags = OCFS2_META_LOCK_RECOVERY | OCFS2_META_LOCK_NOQUEUE;
1796         status = ocfs2_inode_lock_full(inode, NULL, 1, flags);
1797         if (status < 0) {
1798                 if (status != -EAGAIN)
1799                         mlog_errno(status);
1800                 goto bail;
1801         }
1802
1803         ocfs2_inode_unlock(inode, 1);
1804 bail:
1805         iput(inode);
1806
1807         return status;
1808 }
1809
1810 /* Call this underneath ocfs2_super_lock. It also assumes that the
1811  * slot info struct has been updated from disk. */
1812 int ocfs2_mark_dead_nodes(struct ocfs2_super *osb)
1813 {
1814         unsigned int node_num;
1815         int status, i;
1816         u32 gen;
1817         struct buffer_head *bh = NULL;
1818         struct ocfs2_dinode *di;
1819
1820         /* This is called with the super block cluster lock, so we
1821          * know that the slot map can't change underneath us. */
1822
1823         for (i = 0; i < osb->max_slots; i++) {
1824                 /* Read journal inode to get the recovery generation */
1825                 status = ocfs2_read_journal_inode(osb, i, &bh, NULL);
1826                 if (status) {
1827                         mlog_errno(status);
1828                         goto bail;
1829                 }
1830                 di = (struct ocfs2_dinode *)bh->b_data;
1831                 gen = ocfs2_get_recovery_generation(di);
1832                 brelse(bh);
1833                 bh = NULL;
1834
1835                 spin_lock(&osb->osb_lock);
1836                 osb->slot_recovery_generations[i] = gen;
1837
1838                 trace_ocfs2_mark_dead_nodes(i,
1839                                             osb->slot_recovery_generations[i]);
1840
1841                 if (i == osb->slot_num) {
1842                         spin_unlock(&osb->osb_lock);
1843                         continue;
1844                 }
1845
1846                 status = ocfs2_slot_to_node_num_locked(osb, i, &node_num);
1847                 if (status == -ENOENT) {
1848                         spin_unlock(&osb->osb_lock);
1849                         continue;
1850                 }
1851
1852                 if (__ocfs2_recovery_map_test(osb, node_num)) {
1853                         spin_unlock(&osb->osb_lock);
1854                         continue;
1855                 }
1856                 spin_unlock(&osb->osb_lock);
1857
1858                 /* Ok, we have a slot occupied by another node which
1859                  * is not in the recovery map. We trylock his journal
1860                  * file here to test if he's alive. */
1861                 status = ocfs2_trylock_journal(osb, i);
1862                 if (!status) {
1863                         /* Since we're called from mount, we know that
1864                          * the recovery thread can't race us on
1865                          * setting / checking the recovery bits. */
1866                         ocfs2_recovery_thread(osb, node_num);
1867                 } else if ((status < 0) && (status != -EAGAIN)) {
1868                         mlog_errno(status);
1869                         goto bail;
1870                 }
1871         }
1872
1873         status = 0;
1874 bail:
1875         return status;
1876 }
1877
1878 /*
1879  * Scan timer should get fired every ORPHAN_SCAN_SCHEDULE_TIMEOUT. Add some
1880  * randomness to the timeout to minimize multple nodes firing the timer at the
1881  * same time.
1882  */
1883 static inline unsigned long ocfs2_orphan_scan_timeout(void)
1884 {
1885         unsigned long time;
1886
1887         get_random_bytes(&time, sizeof(time));
1888         time = ORPHAN_SCAN_SCHEDULE_TIMEOUT + (time % 5000);
1889         return msecs_to_jiffies(time);
1890 }
1891
1892 /*
1893  * ocfs2_queue_orphan_scan calls ocfs2_queue_recovery_completion for
1894  * every slot, queuing a recovery of the slot on the ocfs2_wq thread. This
1895  * is done to catch any orphans that are left over in orphan directories.
1896  *
1897  * It scans all slots, even ones that are in use. It does so to handle the
1898  * case described below:
1899  *
1900  *   Node 1 has an inode it was using. The dentry went away due to memory
1901  *   pressure.  Node 1 closes the inode, but it's on the free list. The node
1902  *   has the open lock.
1903  *   Node 2 unlinks the inode. It grabs the dentry lock to notify others,
1904  *   but node 1 has no dentry and doesn't get the message. It trylocks the
1905  *   open lock, sees that another node has a PR, and does nothing.
1906  *   Later node 2 runs its orphan dir. It igets the inode, trylocks the
1907  *   open lock, sees the PR still, and does nothing.
1908  *   Basically, we have to trigger an orphan iput on node 1. The only way
1909  *   for this to happen is if node 1 runs node 2's orphan dir.
1910  *
1911  * ocfs2_queue_orphan_scan gets called every ORPHAN_SCAN_SCHEDULE_TIMEOUT
1912  * seconds.  It gets an EX lock on os_lockres and checks sequence number
1913  * stored in LVB. If the sequence number has changed, it means some other
1914  * node has done the scan.  This node skips the scan and tracks the
1915  * sequence number.  If the sequence number didn't change, it means a scan
1916  * hasn't happened.  The node queues a scan and increments the
1917  * sequence number in the LVB.
1918  */
1919 static void ocfs2_queue_orphan_scan(struct ocfs2_super *osb)
1920 {
1921         struct ocfs2_orphan_scan *os;
1922         int status, i;
1923         u32 seqno = 0;
1924
1925         os = &osb->osb_orphan_scan;
1926
1927         if (atomic_read(&os->os_state) == ORPHAN_SCAN_INACTIVE)
1928                 goto out;
1929
1930         trace_ocfs2_queue_orphan_scan_begin(os->os_count, os->os_seqno,
1931                                             atomic_read(&os->os_state));
1932
1933         status = ocfs2_orphan_scan_lock(osb, &seqno);
1934         if (status < 0) {
1935                 if (status != -EAGAIN)
1936                         mlog_errno(status);
1937                 goto out;
1938         }
1939
1940         /* Do no queue the tasks if the volume is being umounted */
1941         if (atomic_read(&os->os_state) == ORPHAN_SCAN_INACTIVE)
1942                 goto unlock;
1943
1944         if (os->os_seqno != seqno) {
1945                 os->os_seqno = seqno;
1946                 goto unlock;
1947         }
1948
1949         for (i = 0; i < osb->max_slots; i++)
1950                 ocfs2_queue_recovery_completion(osb->journal, i, NULL, NULL,
1951                                                 NULL, ORPHAN_NO_NEED_TRUNCATE);
1952         /*
1953          * We queued a recovery on orphan slots, increment the sequence
1954          * number and update LVB so other node will skip the scan for a while
1955          */
1956         seqno++;
1957         os->os_count++;
1958         os->os_scantime = CURRENT_TIME;
1959 unlock:
1960         ocfs2_orphan_scan_unlock(osb, seqno);
1961 out:
1962         trace_ocfs2_queue_orphan_scan_end(os->os_count, os->os_seqno,
1963                                           atomic_read(&os->os_state));
1964         return;
1965 }
1966
1967 /* Worker task that gets fired every ORPHAN_SCAN_SCHEDULE_TIMEOUT millsec */
1968 static void ocfs2_orphan_scan_work(struct work_struct *work)
1969 {
1970         struct ocfs2_orphan_scan *os;
1971         struct ocfs2_super *osb;
1972
1973         os = container_of(work, struct ocfs2_orphan_scan,
1974                           os_orphan_scan_work.work);
1975         osb = os->os_osb;
1976
1977         mutex_lock(&os->os_lock);
1978         ocfs2_queue_orphan_scan(osb);
1979         if (atomic_read(&os->os_state) == ORPHAN_SCAN_ACTIVE)
1980                 queue_delayed_work(osb->ocfs2_wq, &os->os_orphan_scan_work,
1981                                       ocfs2_orphan_scan_timeout());
1982         mutex_unlock(&os->os_lock);
1983 }
1984
1985 void ocfs2_orphan_scan_stop(struct ocfs2_super *osb)
1986 {
1987         struct ocfs2_orphan_scan *os;
1988
1989         os = &osb->osb_orphan_scan;
1990         if (atomic_read(&os->os_state) == ORPHAN_SCAN_ACTIVE) {
1991                 atomic_set(&os->os_state, ORPHAN_SCAN_INACTIVE);
1992                 mutex_lock(&os->os_lock);
1993                 cancel_delayed_work(&os->os_orphan_scan_work);
1994                 mutex_unlock(&os->os_lock);
1995         }
1996 }
1997
1998 void ocfs2_orphan_scan_init(struct ocfs2_super *osb)
1999 {
2000         struct ocfs2_orphan_scan *os;
2001
2002         os = &osb->osb_orphan_scan;
2003         os->os_osb = osb;
2004         os->os_count = 0;
2005         os->os_seqno = 0;
2006         mutex_init(&os->os_lock);
2007         INIT_DELAYED_WORK(&os->os_orphan_scan_work, ocfs2_orphan_scan_work);
2008 }
2009
2010 void ocfs2_orphan_scan_start(struct ocfs2_super *osb)
2011 {
2012         struct ocfs2_orphan_scan *os;
2013
2014         os = &osb->osb_orphan_scan;
2015         os->os_scantime = CURRENT_TIME;
2016         if (ocfs2_is_hard_readonly(osb) || ocfs2_mount_local(osb))
2017                 atomic_set(&os->os_state, ORPHAN_SCAN_INACTIVE);
2018         else {
2019                 atomic_set(&os->os_state, ORPHAN_SCAN_ACTIVE);
2020                 queue_delayed_work(osb->ocfs2_wq, &os->os_orphan_scan_work,
2021                                    ocfs2_orphan_scan_timeout());
2022         }
2023 }
2024
2025 struct ocfs2_orphan_filldir_priv {
2026         struct dir_context      ctx;
2027         struct inode            *head;
2028         struct ocfs2_super      *osb;
2029         enum ocfs2_orphan_reco_type orphan_reco_type;
2030 };
2031
2032 static int ocfs2_orphan_filldir(struct dir_context *ctx, const char *name,
2033                                 int name_len, loff_t pos, u64 ino,
2034                                 unsigned type)
2035 {
2036         struct ocfs2_orphan_filldir_priv *p =
2037                 container_of(ctx, struct ocfs2_orphan_filldir_priv, ctx);
2038         struct inode *iter;
2039
2040         if (name_len == 1 && !strncmp(".", name, 1))
2041                 return 0;
2042         if (name_len == 2 && !strncmp("..", name, 2))
2043                 return 0;
2044
2045         /* do not include dio entry in case of orphan scan */
2046         if ((p->orphan_reco_type == ORPHAN_NO_NEED_TRUNCATE) &&
2047                         (!strncmp(name, OCFS2_DIO_ORPHAN_PREFIX,
2048                         OCFS2_DIO_ORPHAN_PREFIX_LEN)))
2049                 return 0;
2050
2051         /* Skip bad inodes so that recovery can continue */
2052         iter = ocfs2_iget(p->osb, ino,
2053                           OCFS2_FI_FLAG_ORPHAN_RECOVERY, 0);
2054         if (IS_ERR(iter))
2055                 return 0;
2056
2057         if (!strncmp(name, OCFS2_DIO_ORPHAN_PREFIX,
2058                         OCFS2_DIO_ORPHAN_PREFIX_LEN))
2059                 OCFS2_I(iter)->ip_flags |= OCFS2_INODE_DIO_ORPHAN_ENTRY;
2060
2061         /* Skip inodes which are already added to recover list, since dio may
2062          * happen concurrently with unlink/rename */
2063         if (OCFS2_I(iter)->ip_next_orphan) {
2064                 iput(iter);
2065                 return 0;
2066         }
2067
2068         trace_ocfs2_orphan_filldir((unsigned long long)OCFS2_I(iter)->ip_blkno);
2069         /* No locking is required for the next_orphan queue as there
2070          * is only ever a single process doing orphan recovery. */
2071         OCFS2_I(iter)->ip_next_orphan = p->head;
2072         p->head = iter;
2073
2074         return 0;
2075 }
2076
2077 static int ocfs2_queue_orphans(struct ocfs2_super *osb,
2078                                int slot,
2079                                struct inode **head,
2080                                enum ocfs2_orphan_reco_type orphan_reco_type)
2081 {
2082         int status;
2083         struct inode *orphan_dir_inode = NULL;
2084         struct ocfs2_orphan_filldir_priv priv = {
2085                 .ctx.actor = ocfs2_orphan_filldir,
2086                 .osb = osb,
2087                 .head = *head,
2088                 .orphan_reco_type = orphan_reco_type
2089         };
2090
2091         orphan_dir_inode = ocfs2_get_system_file_inode(osb,
2092                                                        ORPHAN_DIR_SYSTEM_INODE,
2093                                                        slot);
2094         if  (!orphan_dir_inode) {
2095                 status = -ENOENT;
2096                 mlog_errno(status);
2097                 return status;
2098         }
2099
2100         inode_lock(orphan_dir_inode);
2101         status = ocfs2_inode_lock(orphan_dir_inode, NULL, 0);
2102         if (status < 0) {
2103                 mlog_errno(status);
2104                 goto out;
2105         }
2106
2107         status = ocfs2_dir_foreach(orphan_dir_inode, &priv.ctx);
2108         if (status) {
2109                 mlog_errno(status);
2110                 goto out_cluster;
2111         }
2112
2113         *head = priv.head;
2114
2115 out_cluster:
2116         ocfs2_inode_unlock(orphan_dir_inode, 0);
2117 out:
2118         inode_unlock(orphan_dir_inode);
2119         iput(orphan_dir_inode);
2120         return status;
2121 }
2122
2123 static int ocfs2_orphan_recovery_can_continue(struct ocfs2_super *osb,
2124                                               int slot)
2125 {
2126         int ret;
2127
2128         spin_lock(&osb->osb_lock);
2129         ret = !osb->osb_orphan_wipes[slot];
2130         spin_unlock(&osb->osb_lock);
2131         return ret;
2132 }
2133
2134 static void ocfs2_mark_recovering_orphan_dir(struct ocfs2_super *osb,
2135                                              int slot)
2136 {
2137         spin_lock(&osb->osb_lock);
2138         /* Mark ourselves such that new processes in delete_inode()
2139          * know to quit early. */
2140         ocfs2_node_map_set_bit(osb, &osb->osb_recovering_orphan_dirs, slot);
2141         while (osb->osb_orphan_wipes[slot]) {
2142                 /* If any processes are already in the middle of an
2143                  * orphan wipe on this dir, then we need to wait for
2144                  * them. */
2145                 spin_unlock(&osb->osb_lock);
2146                 wait_event_interruptible(osb->osb_wipe_event,
2147                                          ocfs2_orphan_recovery_can_continue(osb, slot));
2148                 spin_lock(&osb->osb_lock);
2149         }
2150         spin_unlock(&osb->osb_lock);
2151 }
2152
2153 static void ocfs2_clear_recovering_orphan_dir(struct ocfs2_super *osb,
2154                                               int slot)
2155 {
2156         ocfs2_node_map_clear_bit(osb, &osb->osb_recovering_orphan_dirs, slot);
2157 }
2158
2159 /*
2160  * Orphan recovery. Each mounted node has it's own orphan dir which we
2161  * must run during recovery. Our strategy here is to build a list of
2162  * the inodes in the orphan dir and iget/iput them. The VFS does
2163  * (most) of the rest of the work.
2164  *
2165  * Orphan recovery can happen at any time, not just mount so we have a
2166  * couple of extra considerations.
2167  *
2168  * - We grab as many inodes as we can under the orphan dir lock -
2169  *   doing iget() outside the orphan dir risks getting a reference on
2170  *   an invalid inode.
2171  * - We must be sure not to deadlock with other processes on the
2172  *   system wanting to run delete_inode(). This can happen when they go
2173  *   to lock the orphan dir and the orphan recovery process attempts to
2174  *   iget() inside the orphan dir lock. This can be avoided by
2175  *   advertising our state to ocfs2_delete_inode().
2176  */
2177 static int ocfs2_recover_orphans(struct ocfs2_super *osb,
2178                                  int slot,
2179                                  enum ocfs2_orphan_reco_type orphan_reco_type)
2180 {
2181         int ret = 0;
2182         struct inode *inode = NULL;
2183         struct inode *iter;
2184         struct ocfs2_inode_info *oi;
2185         struct buffer_head *di_bh = NULL;
2186         struct ocfs2_dinode *di = NULL;
2187
2188         trace_ocfs2_recover_orphans(slot);
2189
2190         ocfs2_mark_recovering_orphan_dir(osb, slot);
2191         ret = ocfs2_queue_orphans(osb, slot, &inode, orphan_reco_type);
2192         ocfs2_clear_recovering_orphan_dir(osb, slot);
2193
2194         /* Error here should be noted, but we want to continue with as
2195          * many queued inodes as we've got. */
2196         if (ret)
2197                 mlog_errno(ret);
2198
2199         while (inode) {
2200                 oi = OCFS2_I(inode);
2201                 trace_ocfs2_recover_orphans_iput(
2202                                         (unsigned long long)oi->ip_blkno);
2203
2204                 iter = oi->ip_next_orphan;
2205                 oi->ip_next_orphan = NULL;
2206
2207                 if (oi->ip_flags & OCFS2_INODE_DIO_ORPHAN_ENTRY) {
2208                         inode_lock(inode);
2209                         ret = ocfs2_rw_lock(inode, 1);
2210                         if (ret < 0) {
2211                                 mlog_errno(ret);
2212                                 goto unlock_mutex;
2213                         }
2214                         /*
2215                          * We need to take and drop the inode lock to
2216                          * force read inode from disk.
2217                          */
2218                         ret = ocfs2_inode_lock(inode, &di_bh, 1);
2219                         if (ret) {
2220                                 mlog_errno(ret);
2221                                 goto unlock_rw;
2222                         }
2223
2224                         di = (struct ocfs2_dinode *)di_bh->b_data;
2225
2226                         if (di->i_flags & cpu_to_le32(OCFS2_DIO_ORPHANED_FL)) {
2227                                 ret = ocfs2_truncate_file(inode, di_bh,
2228                                                 i_size_read(inode));
2229                                 if (ret < 0) {
2230                                         if (ret != -ENOSPC)
2231                                                 mlog_errno(ret);
2232                                         goto unlock_inode;
2233                                 }
2234
2235                                 ret = ocfs2_del_inode_from_orphan(osb, inode,
2236                                                 di_bh, 0, 0);
2237                                 if (ret)
2238                                         mlog_errno(ret);
2239                         }
2240 unlock_inode:
2241                         ocfs2_inode_unlock(inode, 1);
2242                         brelse(di_bh);
2243                         di_bh = NULL;
2244 unlock_rw:
2245                         ocfs2_rw_unlock(inode, 1);
2246 unlock_mutex:
2247                         inode_unlock(inode);
2248
2249                         /* clear dio flag in ocfs2_inode_info */
2250                         oi->ip_flags &= ~OCFS2_INODE_DIO_ORPHAN_ENTRY;
2251                 } else {
2252                         spin_lock(&oi->ip_lock);
2253                         /* Set the proper information to get us going into
2254                          * ocfs2_delete_inode. */
2255                         oi->ip_flags |= OCFS2_INODE_MAYBE_ORPHANED;
2256                         spin_unlock(&oi->ip_lock);
2257                 }
2258
2259                 iput(inode);
2260                 inode = iter;
2261         }
2262
2263         return ret;
2264 }
2265
2266 static int __ocfs2_wait_on_mount(struct ocfs2_super *osb, int quota)
2267 {
2268         /* This check is good because ocfs2 will wait on our recovery
2269          * thread before changing it to something other than MOUNTED
2270          * or DISABLED. */
2271         wait_event(osb->osb_mount_event,
2272                   (!quota && atomic_read(&osb->vol_state) == VOLUME_MOUNTED) ||
2273                    atomic_read(&osb->vol_state) == VOLUME_MOUNTED_QUOTAS ||
2274                    atomic_read(&osb->vol_state) == VOLUME_DISABLED);
2275
2276         /* If there's an error on mount, then we may never get to the
2277          * MOUNTED flag, but this is set right before
2278          * dismount_volume() so we can trust it. */
2279         if (atomic_read(&osb->vol_state) == VOLUME_DISABLED) {
2280                 trace_ocfs2_wait_on_mount(VOLUME_DISABLED);
2281                 mlog(0, "mount error, exiting!\n");
2282                 return -EBUSY;
2283         }
2284
2285         return 0;
2286 }
2287
2288 static int ocfs2_commit_thread(void *arg)
2289 {
2290         int status;
2291         struct ocfs2_super *osb = arg;
2292         struct ocfs2_journal *journal = osb->journal;
2293
2294         /* we can trust j_num_trans here because _should_stop() is only set in
2295          * shutdown and nobody other than ourselves should be able to start
2296          * transactions.  committing on shutdown might take a few iterations
2297          * as final transactions put deleted inodes on the list */
2298         while (!(kthread_should_stop() &&
2299                  atomic_read(&journal->j_num_trans) == 0)) {
2300
2301                 wait_event_interruptible(osb->checkpoint_event,
2302                                          atomic_read(&journal->j_num_trans)
2303                                          || kthread_should_stop());
2304
2305                 status = ocfs2_commit_cache(osb);
2306                 if (status < 0) {
2307                         static unsigned long abort_warn_time;
2308
2309                         /* Warn about this once per minute */
2310                         if (printk_timed_ratelimit(&abort_warn_time, 60*HZ))
2311                                 mlog(ML_ERROR, "status = %d, journal is "
2312                                                 "already aborted.\n", status);
2313                         /*
2314                          * After ocfs2_commit_cache() fails, j_num_trans has a
2315                          * non-zero value.  Sleep here to avoid a busy-wait
2316                          * loop.
2317                          */
2318                         msleep_interruptible(1000);
2319                 }
2320
2321                 if (kthread_should_stop() && atomic_read(&journal->j_num_trans)){
2322                         mlog(ML_KTHREAD,
2323                              "commit_thread: %u transactions pending on "
2324                              "shutdown\n",
2325                              atomic_read(&journal->j_num_trans));
2326                 }
2327         }
2328
2329         return 0;
2330 }
2331
2332 /* Reads all the journal inodes without taking any cluster locks. Used
2333  * for hard readonly access to determine whether any journal requires
2334  * recovery. Also used to refresh the recovery generation numbers after
2335  * a journal has been recovered by another node.
2336  */
2337 int ocfs2_check_journals_nolocks(struct ocfs2_super *osb)
2338 {
2339         int ret = 0;
2340         unsigned int slot;
2341         struct buffer_head *di_bh = NULL;
2342         struct ocfs2_dinode *di;
2343         int journal_dirty = 0;
2344
2345         for(slot = 0; slot < osb->max_slots; slot++) {
2346                 ret = ocfs2_read_journal_inode(osb, slot, &di_bh, NULL);
2347                 if (ret) {
2348                         mlog_errno(ret);
2349                         goto out;
2350                 }
2351
2352                 di = (struct ocfs2_dinode *) di_bh->b_data;
2353
2354                 osb->slot_recovery_generations[slot] =
2355                                         ocfs2_get_recovery_generation(di);
2356
2357                 if (le32_to_cpu(di->id1.journal1.ij_flags) &
2358                     OCFS2_JOURNAL_DIRTY_FL)
2359                         journal_dirty = 1;
2360
2361                 brelse(di_bh);
2362                 di_bh = NULL;
2363         }
2364
2365 out:
2366         if (journal_dirty)
2367                 ret = -EROFS;
2368         return ret;
2369 }