GNU Linux-libre 4.9.308-gnu1
[releases.git] / fs / jbd2 / journal.c
1 /*
2  * linux/fs/jbd2/journal.c
3  *
4  * Written by Stephen C. Tweedie <sct@redhat.com>, 1998
5  *
6  * Copyright 1998 Red Hat corp --- All Rights Reserved
7  *
8  * This file is part of the Linux kernel and is made available under
9  * the terms of the GNU General Public License, version 2, or at your
10  * option, any later version, incorporated herein by reference.
11  *
12  * Generic filesystem journal-writing code; part of the ext2fs
13  * journaling system.
14  *
15  * This file manages journals: areas of disk reserved for logging
16  * transactional updates.  This includes the kernel journaling thread
17  * which is responsible for scheduling updates to the log.
18  *
19  * We do not actually manage the physical storage of the journal in this
20  * file: that is left to a per-journal policy function, which allows us
21  * to store the journal within a filesystem-specified area for ext2
22  * journaling (ext2 can use a reserved inode for storing the log).
23  */
24
25 #include <linux/module.h>
26 #include <linux/time.h>
27 #include <linux/fs.h>
28 #include <linux/jbd2.h>
29 #include <linux/errno.h>
30 #include <linux/slab.h>
31 #include <linux/init.h>
32 #include <linux/mm.h>
33 #include <linux/freezer.h>
34 #include <linux/pagemap.h>
35 #include <linux/kthread.h>
36 #include <linux/poison.h>
37 #include <linux/proc_fs.h>
38 #include <linux/seq_file.h>
39 #include <linux/math64.h>
40 #include <linux/hash.h>
41 #include <linux/log2.h>
42 #include <linux/vmalloc.h>
43 #include <linux/backing-dev.h>
44 #include <linux/bitops.h>
45 #include <linux/ratelimit.h>
46
47 #define CREATE_TRACE_POINTS
48 #include <trace/events/jbd2.h>
49
50 #include <asm/uaccess.h>
51 #include <asm/page.h>
52
53 #ifdef CONFIG_JBD2_DEBUG
54 ushort jbd2_journal_enable_debug __read_mostly;
55 EXPORT_SYMBOL(jbd2_journal_enable_debug);
56
57 module_param_named(jbd2_debug, jbd2_journal_enable_debug, ushort, 0644);
58 MODULE_PARM_DESC(jbd2_debug, "Debugging level for jbd2");
59 #endif
60
61 EXPORT_SYMBOL(jbd2_journal_extend);
62 EXPORT_SYMBOL(jbd2_journal_stop);
63 EXPORT_SYMBOL(jbd2_journal_lock_updates);
64 EXPORT_SYMBOL(jbd2_journal_unlock_updates);
65 EXPORT_SYMBOL(jbd2_journal_get_write_access);
66 EXPORT_SYMBOL(jbd2_journal_get_create_access);
67 EXPORT_SYMBOL(jbd2_journal_get_undo_access);
68 EXPORT_SYMBOL(jbd2_journal_set_triggers);
69 EXPORT_SYMBOL(jbd2_journal_dirty_metadata);
70 EXPORT_SYMBOL(jbd2_journal_forget);
71 #if 0
72 EXPORT_SYMBOL(journal_sync_buffer);
73 #endif
74 EXPORT_SYMBOL(jbd2_journal_flush);
75 EXPORT_SYMBOL(jbd2_journal_revoke);
76
77 EXPORT_SYMBOL(jbd2_journal_init_dev);
78 EXPORT_SYMBOL(jbd2_journal_init_inode);
79 EXPORT_SYMBOL(jbd2_journal_check_used_features);
80 EXPORT_SYMBOL(jbd2_journal_check_available_features);
81 EXPORT_SYMBOL(jbd2_journal_set_features);
82 EXPORT_SYMBOL(jbd2_journal_load);
83 EXPORT_SYMBOL(jbd2_journal_destroy);
84 EXPORT_SYMBOL(jbd2_journal_abort);
85 EXPORT_SYMBOL(jbd2_journal_errno);
86 EXPORT_SYMBOL(jbd2_journal_ack_err);
87 EXPORT_SYMBOL(jbd2_journal_clear_err);
88 EXPORT_SYMBOL(jbd2_log_wait_commit);
89 EXPORT_SYMBOL(jbd2_log_start_commit);
90 EXPORT_SYMBOL(jbd2_journal_start_commit);
91 EXPORT_SYMBOL(jbd2_journal_force_commit_nested);
92 EXPORT_SYMBOL(jbd2_journal_wipe);
93 EXPORT_SYMBOL(jbd2_journal_blocks_per_page);
94 EXPORT_SYMBOL(jbd2_journal_invalidatepage);
95 EXPORT_SYMBOL(jbd2_journal_try_to_free_buffers);
96 EXPORT_SYMBOL(jbd2_journal_force_commit);
97 EXPORT_SYMBOL(jbd2_journal_inode_add_write);
98 EXPORT_SYMBOL(jbd2_journal_inode_add_wait);
99 EXPORT_SYMBOL(jbd2_journal_init_jbd_inode);
100 EXPORT_SYMBOL(jbd2_journal_release_jbd_inode);
101 EXPORT_SYMBOL(jbd2_journal_begin_ordered_truncate);
102 EXPORT_SYMBOL(jbd2_inode_cache);
103
104 static void __journal_abort_soft (journal_t *journal, int errno);
105 static int jbd2_journal_create_slab(size_t slab_size);
106
107 #ifdef CONFIG_JBD2_DEBUG
108 void __jbd2_debug(int level, const char *file, const char *func,
109                   unsigned int line, const char *fmt, ...)
110 {
111         struct va_format vaf;
112         va_list args;
113
114         if (level > jbd2_journal_enable_debug)
115                 return;
116         va_start(args, fmt);
117         vaf.fmt = fmt;
118         vaf.va = &args;
119         printk(KERN_DEBUG "%s: (%s, %u): %pV\n", file, func, line, &vaf);
120         va_end(args);
121 }
122 EXPORT_SYMBOL(__jbd2_debug);
123 #endif
124
125 /* Checksumming functions */
126 static int jbd2_verify_csum_type(journal_t *j, journal_superblock_t *sb)
127 {
128         if (!jbd2_journal_has_csum_v2or3_feature(j))
129                 return 1;
130
131         return sb->s_checksum_type == JBD2_CRC32C_CHKSUM;
132 }
133
134 static __be32 jbd2_superblock_csum(journal_t *j, journal_superblock_t *sb)
135 {
136         __u32 csum;
137         __be32 old_csum;
138
139         old_csum = sb->s_checksum;
140         sb->s_checksum = 0;
141         csum = jbd2_chksum(j, ~0, (char *)sb, sizeof(journal_superblock_t));
142         sb->s_checksum = old_csum;
143
144         return cpu_to_be32(csum);
145 }
146
147 static int jbd2_superblock_csum_verify(journal_t *j, journal_superblock_t *sb)
148 {
149         if (!jbd2_journal_has_csum_v2or3(j))
150                 return 1;
151
152         return sb->s_checksum == jbd2_superblock_csum(j, sb);
153 }
154
155 static void jbd2_superblock_csum_set(journal_t *j, journal_superblock_t *sb)
156 {
157         if (!jbd2_journal_has_csum_v2or3(j))
158                 return;
159
160         sb->s_checksum = jbd2_superblock_csum(j, sb);
161 }
162
163 /*
164  * Helper function used to manage commit timeouts
165  */
166
167 static void commit_timeout(unsigned long __data)
168 {
169         struct task_struct * p = (struct task_struct *) __data;
170
171         wake_up_process(p);
172 }
173
174 /*
175  * kjournald2: The main thread function used to manage a logging device
176  * journal.
177  *
178  * This kernel thread is responsible for two things:
179  *
180  * 1) COMMIT:  Every so often we need to commit the current state of the
181  *    filesystem to disk.  The journal thread is responsible for writing
182  *    all of the metadata buffers to disk.
183  *
184  * 2) CHECKPOINT: We cannot reuse a used section of the log file until all
185  *    of the data in that part of the log has been rewritten elsewhere on
186  *    the disk.  Flushing these old buffers to reclaim space in the log is
187  *    known as checkpointing, and this thread is responsible for that job.
188  */
189
190 static int kjournald2(void *arg)
191 {
192         journal_t *journal = arg;
193         transaction_t *transaction;
194
195         /*
196          * Set up an interval timer which can be used to trigger a commit wakeup
197          * after the commit interval expires
198          */
199         setup_timer(&journal->j_commit_timer, commit_timeout,
200                         (unsigned long)current);
201
202         set_freezable();
203
204         /* Record that the journal thread is running */
205         journal->j_task = current;
206         wake_up(&journal->j_wait_done_commit);
207
208         /*
209          * And now, wait forever for commit wakeup events.
210          */
211         write_lock(&journal->j_state_lock);
212
213 loop:
214         if (journal->j_flags & JBD2_UNMOUNT)
215                 goto end_loop;
216
217         jbd_debug(1, "commit_sequence=%d, commit_request=%d\n",
218                 journal->j_commit_sequence, journal->j_commit_request);
219
220         if (journal->j_commit_sequence != journal->j_commit_request) {
221                 jbd_debug(1, "OK, requests differ\n");
222                 write_unlock(&journal->j_state_lock);
223                 del_timer_sync(&journal->j_commit_timer);
224                 jbd2_journal_commit_transaction(journal);
225                 write_lock(&journal->j_state_lock);
226                 goto loop;
227         }
228
229         wake_up(&journal->j_wait_done_commit);
230         if (freezing(current)) {
231                 /*
232                  * The simpler the better. Flushing journal isn't a
233                  * good idea, because that depends on threads that may
234                  * be already stopped.
235                  */
236                 jbd_debug(1, "Now suspending kjournald2\n");
237                 write_unlock(&journal->j_state_lock);
238                 try_to_freeze();
239                 write_lock(&journal->j_state_lock);
240         } else {
241                 /*
242                  * We assume on resume that commits are already there,
243                  * so we don't sleep
244                  */
245                 DEFINE_WAIT(wait);
246                 int should_sleep = 1;
247
248                 prepare_to_wait(&journal->j_wait_commit, &wait,
249                                 TASK_INTERRUPTIBLE);
250                 if (journal->j_commit_sequence != journal->j_commit_request)
251                         should_sleep = 0;
252                 transaction = journal->j_running_transaction;
253                 if (transaction && time_after_eq(jiffies,
254                                                 transaction->t_expires))
255                         should_sleep = 0;
256                 if (journal->j_flags & JBD2_UNMOUNT)
257                         should_sleep = 0;
258                 if (should_sleep) {
259                         write_unlock(&journal->j_state_lock);
260                         schedule();
261                         write_lock(&journal->j_state_lock);
262                 }
263                 finish_wait(&journal->j_wait_commit, &wait);
264         }
265
266         jbd_debug(1, "kjournald2 wakes\n");
267
268         /*
269          * Were we woken up by a commit wakeup event?
270          */
271         transaction = journal->j_running_transaction;
272         if (transaction && time_after_eq(jiffies, transaction->t_expires)) {
273                 journal->j_commit_request = transaction->t_tid;
274                 jbd_debug(1, "woke because of timeout\n");
275         }
276         goto loop;
277
278 end_loop:
279         del_timer_sync(&journal->j_commit_timer);
280         journal->j_task = NULL;
281         wake_up(&journal->j_wait_done_commit);
282         jbd_debug(1, "Journal thread exiting.\n");
283         write_unlock(&journal->j_state_lock);
284         return 0;
285 }
286
287 static int jbd2_journal_start_thread(journal_t *journal)
288 {
289         struct task_struct *t;
290
291         t = kthread_run(kjournald2, journal, "jbd2/%s",
292                         journal->j_devname);
293         if (IS_ERR(t))
294                 return PTR_ERR(t);
295
296         wait_event(journal->j_wait_done_commit, journal->j_task != NULL);
297         return 0;
298 }
299
300 static void journal_kill_thread(journal_t *journal)
301 {
302         write_lock(&journal->j_state_lock);
303         journal->j_flags |= JBD2_UNMOUNT;
304
305         while (journal->j_task) {
306                 write_unlock(&journal->j_state_lock);
307                 wake_up(&journal->j_wait_commit);
308                 wait_event(journal->j_wait_done_commit, journal->j_task == NULL);
309                 write_lock(&journal->j_state_lock);
310         }
311         write_unlock(&journal->j_state_lock);
312 }
313
314 /*
315  * jbd2_journal_write_metadata_buffer: write a metadata buffer to the journal.
316  *
317  * Writes a metadata buffer to a given disk block.  The actual IO is not
318  * performed but a new buffer_head is constructed which labels the data
319  * to be written with the correct destination disk block.
320  *
321  * Any magic-number escaping which needs to be done will cause a
322  * copy-out here.  If the buffer happens to start with the
323  * JBD2_MAGIC_NUMBER, then we can't write it to the log directly: the
324  * magic number is only written to the log for descripter blocks.  In
325  * this case, we copy the data and replace the first word with 0, and we
326  * return a result code which indicates that this buffer needs to be
327  * marked as an escaped buffer in the corresponding log descriptor
328  * block.  The missing word can then be restored when the block is read
329  * during recovery.
330  *
331  * If the source buffer has already been modified by a new transaction
332  * since we took the last commit snapshot, we use the frozen copy of
333  * that data for IO. If we end up using the existing buffer_head's data
334  * for the write, then we have to make sure nobody modifies it while the
335  * IO is in progress. do_get_write_access() handles this.
336  *
337  * The function returns a pointer to the buffer_head to be used for IO.
338  * 
339  *
340  * Return value:
341  *  <0: Error
342  * >=0: Finished OK
343  *
344  * On success:
345  * Bit 0 set == escape performed on the data
346  * Bit 1 set == buffer copy-out performed (kfree the data after IO)
347  */
348
349 int jbd2_journal_write_metadata_buffer(transaction_t *transaction,
350                                   struct journal_head  *jh_in,
351                                   struct buffer_head **bh_out,
352                                   sector_t blocknr)
353 {
354         int need_copy_out = 0;
355         int done_copy_out = 0;
356         int do_escape = 0;
357         char *mapped_data;
358         struct buffer_head *new_bh;
359         struct page *new_page;
360         unsigned int new_offset;
361         struct buffer_head *bh_in = jh2bh(jh_in);
362         journal_t *journal = transaction->t_journal;
363
364         /*
365          * The buffer really shouldn't be locked: only the current committing
366          * transaction is allowed to write it, so nobody else is allowed
367          * to do any IO.
368          *
369          * akpm: except if we're journalling data, and write() output is
370          * also part of a shared mapping, and another thread has
371          * decided to launch a writepage() against this buffer.
372          */
373         J_ASSERT_BH(bh_in, buffer_jbddirty(bh_in));
374
375         new_bh = alloc_buffer_head(GFP_NOFS|__GFP_NOFAIL);
376
377         /* keep subsequent assertions sane */
378         atomic_set(&new_bh->b_count, 1);
379
380         jbd_lock_bh_state(bh_in);
381 repeat:
382         /*
383          * If a new transaction has already done a buffer copy-out, then
384          * we use that version of the data for the commit.
385          */
386         if (jh_in->b_frozen_data) {
387                 done_copy_out = 1;
388                 new_page = virt_to_page(jh_in->b_frozen_data);
389                 new_offset = offset_in_page(jh_in->b_frozen_data);
390         } else {
391                 new_page = jh2bh(jh_in)->b_page;
392                 new_offset = offset_in_page(jh2bh(jh_in)->b_data);
393         }
394
395         mapped_data = kmap_atomic(new_page);
396         /*
397          * Fire data frozen trigger if data already wasn't frozen.  Do this
398          * before checking for escaping, as the trigger may modify the magic
399          * offset.  If a copy-out happens afterwards, it will have the correct
400          * data in the buffer.
401          */
402         if (!done_copy_out)
403                 jbd2_buffer_frozen_trigger(jh_in, mapped_data + new_offset,
404                                            jh_in->b_triggers);
405
406         /*
407          * Check for escaping
408          */
409         if (*((__be32 *)(mapped_data + new_offset)) ==
410                                 cpu_to_be32(JBD2_MAGIC_NUMBER)) {
411                 need_copy_out = 1;
412                 do_escape = 1;
413         }
414         kunmap_atomic(mapped_data);
415
416         /*
417          * Do we need to do a data copy?
418          */
419         if (need_copy_out && !done_copy_out) {
420                 char *tmp;
421
422                 jbd_unlock_bh_state(bh_in);
423                 tmp = jbd2_alloc(bh_in->b_size, GFP_NOFS);
424                 if (!tmp) {
425                         brelse(new_bh);
426                         return -ENOMEM;
427                 }
428                 jbd_lock_bh_state(bh_in);
429                 if (jh_in->b_frozen_data) {
430                         jbd2_free(tmp, bh_in->b_size);
431                         goto repeat;
432                 }
433
434                 jh_in->b_frozen_data = tmp;
435                 mapped_data = kmap_atomic(new_page);
436                 memcpy(tmp, mapped_data + new_offset, bh_in->b_size);
437                 kunmap_atomic(mapped_data);
438
439                 new_page = virt_to_page(tmp);
440                 new_offset = offset_in_page(tmp);
441                 done_copy_out = 1;
442
443                 /*
444                  * This isn't strictly necessary, as we're using frozen
445                  * data for the escaping, but it keeps consistency with
446                  * b_frozen_data usage.
447                  */
448                 jh_in->b_frozen_triggers = jh_in->b_triggers;
449         }
450
451         /*
452          * Did we need to do an escaping?  Now we've done all the
453          * copying, we can finally do so.
454          */
455         if (do_escape) {
456                 mapped_data = kmap_atomic(new_page);
457                 *((unsigned int *)(mapped_data + new_offset)) = 0;
458                 kunmap_atomic(mapped_data);
459         }
460
461         set_bh_page(new_bh, new_page, new_offset);
462         new_bh->b_size = bh_in->b_size;
463         new_bh->b_bdev = journal->j_dev;
464         new_bh->b_blocknr = blocknr;
465         new_bh->b_private = bh_in;
466         set_buffer_mapped(new_bh);
467         set_buffer_dirty(new_bh);
468
469         *bh_out = new_bh;
470
471         /*
472          * The to-be-written buffer needs to get moved to the io queue,
473          * and the original buffer whose contents we are shadowing or
474          * copying is moved to the transaction's shadow queue.
475          */
476         JBUFFER_TRACE(jh_in, "file as BJ_Shadow");
477         spin_lock(&journal->j_list_lock);
478         __jbd2_journal_file_buffer(jh_in, transaction, BJ_Shadow);
479         spin_unlock(&journal->j_list_lock);
480         set_buffer_shadow(bh_in);
481         jbd_unlock_bh_state(bh_in);
482
483         return do_escape | (done_copy_out << 1);
484 }
485
486 /*
487  * Allocation code for the journal file.  Manage the space left in the
488  * journal, so that we can begin checkpointing when appropriate.
489  */
490
491 /*
492  * Called with j_state_lock locked for writing.
493  * Returns true if a transaction commit was started.
494  */
495 int __jbd2_log_start_commit(journal_t *journal, tid_t target)
496 {
497         /* Return if the txn has already requested to be committed */
498         if (journal->j_commit_request == target)
499                 return 0;
500
501         /*
502          * The only transaction we can possibly wait upon is the
503          * currently running transaction (if it exists).  Otherwise,
504          * the target tid must be an old one.
505          */
506         if (journal->j_running_transaction &&
507             journal->j_running_transaction->t_tid == target) {
508                 /*
509                  * We want a new commit: OK, mark the request and wakeup the
510                  * commit thread.  We do _not_ do the commit ourselves.
511                  */
512
513                 journal->j_commit_request = target;
514                 jbd_debug(1, "JBD2: requesting commit %d/%d\n",
515                           journal->j_commit_request,
516                           journal->j_commit_sequence);
517                 journal->j_running_transaction->t_requested = jiffies;
518                 wake_up(&journal->j_wait_commit);
519                 return 1;
520         } else if (!tid_geq(journal->j_commit_request, target))
521                 /* This should never happen, but if it does, preserve
522                    the evidence before kjournald goes into a loop and
523                    increments j_commit_sequence beyond all recognition. */
524                 WARN_ONCE(1, "JBD2: bad log_start_commit: %u %u %u %u\n",
525                           journal->j_commit_request,
526                           journal->j_commit_sequence,
527                           target, journal->j_running_transaction ? 
528                           journal->j_running_transaction->t_tid : 0);
529         return 0;
530 }
531
532 int jbd2_log_start_commit(journal_t *journal, tid_t tid)
533 {
534         int ret;
535
536         write_lock(&journal->j_state_lock);
537         ret = __jbd2_log_start_commit(journal, tid);
538         write_unlock(&journal->j_state_lock);
539         return ret;
540 }
541
542 /*
543  * Force and wait any uncommitted transactions.  We can only force the running
544  * transaction if we don't have an active handle, otherwise, we will deadlock.
545  * Returns: <0 in case of error,
546  *           0 if nothing to commit,
547  *           1 if transaction was successfully committed.
548  */
549 static int __jbd2_journal_force_commit(journal_t *journal)
550 {
551         transaction_t *transaction = NULL;
552         tid_t tid;
553         int need_to_start = 0, ret = 0;
554
555         read_lock(&journal->j_state_lock);
556         if (journal->j_running_transaction && !current->journal_info) {
557                 transaction = journal->j_running_transaction;
558                 if (!tid_geq(journal->j_commit_request, transaction->t_tid))
559                         need_to_start = 1;
560         } else if (journal->j_committing_transaction)
561                 transaction = journal->j_committing_transaction;
562
563         if (!transaction) {
564                 /* Nothing to commit */
565                 read_unlock(&journal->j_state_lock);
566                 return 0;
567         }
568         tid = transaction->t_tid;
569         read_unlock(&journal->j_state_lock);
570         if (need_to_start)
571                 jbd2_log_start_commit(journal, tid);
572         ret = jbd2_log_wait_commit(journal, tid);
573         if (!ret)
574                 ret = 1;
575
576         return ret;
577 }
578
579 /**
580  * Force and wait upon a commit if the calling process is not within
581  * transaction.  This is used for forcing out undo-protected data which contains
582  * bitmaps, when the fs is running out of space.
583  *
584  * @journal: journal to force
585  * Returns true if progress was made.
586  */
587 int jbd2_journal_force_commit_nested(journal_t *journal)
588 {
589         int ret;
590
591         ret = __jbd2_journal_force_commit(journal);
592         return ret > 0;
593 }
594
595 /**
596  * int journal_force_commit() - force any uncommitted transactions
597  * @journal: journal to force
598  *
599  * Caller want unconditional commit. We can only force the running transaction
600  * if we don't have an active handle, otherwise, we will deadlock.
601  */
602 int jbd2_journal_force_commit(journal_t *journal)
603 {
604         int ret;
605
606         J_ASSERT(!current->journal_info);
607         ret = __jbd2_journal_force_commit(journal);
608         if (ret > 0)
609                 ret = 0;
610         return ret;
611 }
612
613 /*
614  * Start a commit of the current running transaction (if any).  Returns true
615  * if a transaction is going to be committed (or is currently already
616  * committing), and fills its tid in at *ptid
617  */
618 int jbd2_journal_start_commit(journal_t *journal, tid_t *ptid)
619 {
620         int ret = 0;
621
622         write_lock(&journal->j_state_lock);
623         if (journal->j_running_transaction) {
624                 tid_t tid = journal->j_running_transaction->t_tid;
625
626                 __jbd2_log_start_commit(journal, tid);
627                 /* There's a running transaction and we've just made sure
628                  * it's commit has been scheduled. */
629                 if (ptid)
630                         *ptid = tid;
631                 ret = 1;
632         } else if (journal->j_committing_transaction) {
633                 /*
634                  * If commit has been started, then we have to wait for
635                  * completion of that transaction.
636                  */
637                 if (ptid)
638                         *ptid = journal->j_committing_transaction->t_tid;
639                 ret = 1;
640         }
641         write_unlock(&journal->j_state_lock);
642         return ret;
643 }
644
645 /*
646  * Return 1 if a given transaction has not yet sent barrier request
647  * connected with a transaction commit. If 0 is returned, transaction
648  * may or may not have sent the barrier. Used to avoid sending barrier
649  * twice in common cases.
650  */
651 int jbd2_trans_will_send_data_barrier(journal_t *journal, tid_t tid)
652 {
653         int ret = 0;
654         transaction_t *commit_trans;
655
656         if (!(journal->j_flags & JBD2_BARRIER))
657                 return 0;
658         read_lock(&journal->j_state_lock);
659         /* Transaction already committed? */
660         if (tid_geq(journal->j_commit_sequence, tid))
661                 goto out;
662         commit_trans = journal->j_committing_transaction;
663         if (!commit_trans || commit_trans->t_tid != tid) {
664                 ret = 1;
665                 goto out;
666         }
667         /*
668          * Transaction is being committed and we already proceeded to
669          * submitting a flush to fs partition?
670          */
671         if (journal->j_fs_dev != journal->j_dev) {
672                 if (!commit_trans->t_need_data_flush ||
673                     commit_trans->t_state >= T_COMMIT_DFLUSH)
674                         goto out;
675         } else {
676                 if (commit_trans->t_state >= T_COMMIT_JFLUSH)
677                         goto out;
678         }
679         ret = 1;
680 out:
681         read_unlock(&journal->j_state_lock);
682         return ret;
683 }
684 EXPORT_SYMBOL(jbd2_trans_will_send_data_barrier);
685
686 /*
687  * Wait for a specified commit to complete.
688  * The caller may not hold the journal lock.
689  */
690 int jbd2_log_wait_commit(journal_t *journal, tid_t tid)
691 {
692         int err = 0;
693
694         read_lock(&journal->j_state_lock);
695 #ifdef CONFIG_PROVE_LOCKING
696         /*
697          * Some callers make sure transaction is already committing and in that
698          * case we cannot block on open handles anymore. So don't warn in that
699          * case.
700          */
701         if (tid_gt(tid, journal->j_commit_sequence) &&
702             (!journal->j_committing_transaction ||
703              journal->j_committing_transaction->t_tid != tid)) {
704                 read_unlock(&journal->j_state_lock);
705                 jbd2_might_wait_for_commit(journal);
706                 read_lock(&journal->j_state_lock);
707         }
708 #endif
709 #ifdef CONFIG_JBD2_DEBUG
710         if (!tid_geq(journal->j_commit_request, tid)) {
711                 printk(KERN_ERR
712                        "%s: error: j_commit_request=%d, tid=%d\n",
713                        __func__, journal->j_commit_request, tid);
714         }
715 #endif
716         while (tid_gt(tid, journal->j_commit_sequence)) {
717                 jbd_debug(1, "JBD2: want %d, j_commit_sequence=%d\n",
718                                   tid, journal->j_commit_sequence);
719                 read_unlock(&journal->j_state_lock);
720                 wake_up(&journal->j_wait_commit);
721                 wait_event(journal->j_wait_done_commit,
722                                 !tid_gt(tid, journal->j_commit_sequence));
723                 read_lock(&journal->j_state_lock);
724         }
725         read_unlock(&journal->j_state_lock);
726
727         if (unlikely(is_journal_aborted(journal)))
728                 err = -EIO;
729         return err;
730 }
731
732 /*
733  * When this function returns the transaction corresponding to tid
734  * will be completed.  If the transaction has currently running, start
735  * committing that transaction before waiting for it to complete.  If
736  * the transaction id is stale, it is by definition already completed,
737  * so just return SUCCESS.
738  */
739 int jbd2_complete_transaction(journal_t *journal, tid_t tid)
740 {
741         int     need_to_wait = 1;
742
743         read_lock(&journal->j_state_lock);
744         if (journal->j_running_transaction &&
745             journal->j_running_transaction->t_tid == tid) {
746                 if (journal->j_commit_request != tid) {
747                         /* transaction not yet started, so request it */
748                         read_unlock(&journal->j_state_lock);
749                         jbd2_log_start_commit(journal, tid);
750                         goto wait_commit;
751                 }
752         } else if (!(journal->j_committing_transaction &&
753                      journal->j_committing_transaction->t_tid == tid))
754                 need_to_wait = 0;
755         read_unlock(&journal->j_state_lock);
756         if (!need_to_wait)
757                 return 0;
758 wait_commit:
759         return jbd2_log_wait_commit(journal, tid);
760 }
761 EXPORT_SYMBOL(jbd2_complete_transaction);
762
763 /*
764  * Log buffer allocation routines:
765  */
766
767 int jbd2_journal_next_log_block(journal_t *journal, unsigned long long *retp)
768 {
769         unsigned long blocknr;
770
771         write_lock(&journal->j_state_lock);
772         J_ASSERT(journal->j_free > 1);
773
774         blocknr = journal->j_head;
775         journal->j_head++;
776         journal->j_free--;
777         if (journal->j_head == journal->j_last)
778                 journal->j_head = journal->j_first;
779         write_unlock(&journal->j_state_lock);
780         return jbd2_journal_bmap(journal, blocknr, retp);
781 }
782
783 /*
784  * Conversion of logical to physical block numbers for the journal
785  *
786  * On external journals the journal blocks are identity-mapped, so
787  * this is a no-op.  If needed, we can use j_blk_offset - everything is
788  * ready.
789  */
790 int jbd2_journal_bmap(journal_t *journal, unsigned long blocknr,
791                  unsigned long long *retp)
792 {
793         int err = 0;
794         unsigned long long ret;
795
796         if (journal->j_inode) {
797                 ret = bmap(journal->j_inode, blocknr);
798                 if (ret)
799                         *retp = ret;
800                 else {
801                         printk(KERN_ALERT "%s: journal block not found "
802                                         "at offset %lu on %s\n",
803                                __func__, blocknr, journal->j_devname);
804                         err = -EIO;
805                         __journal_abort_soft(journal, err);
806                 }
807         } else {
808                 *retp = blocknr; /* +journal->j_blk_offset */
809         }
810         return err;
811 }
812
813 /*
814  * We play buffer_head aliasing tricks to write data/metadata blocks to
815  * the journal without copying their contents, but for journal
816  * descriptor blocks we do need to generate bona fide buffers.
817  *
818  * After the caller of jbd2_journal_get_descriptor_buffer() has finished modifying
819  * the buffer's contents they really should run flush_dcache_page(bh->b_page).
820  * But we don't bother doing that, so there will be coherency problems with
821  * mmaps of blockdevs which hold live JBD-controlled filesystems.
822  */
823 struct buffer_head *
824 jbd2_journal_get_descriptor_buffer(transaction_t *transaction, int type)
825 {
826         journal_t *journal = transaction->t_journal;
827         struct buffer_head *bh;
828         unsigned long long blocknr;
829         journal_header_t *header;
830         int err;
831
832         err = jbd2_journal_next_log_block(journal, &blocknr);
833
834         if (err)
835                 return NULL;
836
837         bh = __getblk(journal->j_dev, blocknr, journal->j_blocksize);
838         if (!bh)
839                 return NULL;
840         lock_buffer(bh);
841         memset(bh->b_data, 0, journal->j_blocksize);
842         header = (journal_header_t *)bh->b_data;
843         header->h_magic = cpu_to_be32(JBD2_MAGIC_NUMBER);
844         header->h_blocktype = cpu_to_be32(type);
845         header->h_sequence = cpu_to_be32(transaction->t_tid);
846         set_buffer_uptodate(bh);
847         unlock_buffer(bh);
848         BUFFER_TRACE(bh, "return this buffer");
849         return bh;
850 }
851
852 void jbd2_descriptor_block_csum_set(journal_t *j, struct buffer_head *bh)
853 {
854         struct jbd2_journal_block_tail *tail;
855         __u32 csum;
856
857         if (!jbd2_journal_has_csum_v2or3(j))
858                 return;
859
860         tail = (struct jbd2_journal_block_tail *)(bh->b_data + j->j_blocksize -
861                         sizeof(struct jbd2_journal_block_tail));
862         tail->t_checksum = 0;
863         csum = jbd2_chksum(j, j->j_csum_seed, bh->b_data, j->j_blocksize);
864         tail->t_checksum = cpu_to_be32(csum);
865 }
866
867 /*
868  * Return tid of the oldest transaction in the journal and block in the journal
869  * where the transaction starts.
870  *
871  * If the journal is now empty, return which will be the next transaction ID
872  * we will write and where will that transaction start.
873  *
874  * The return value is 0 if journal tail cannot be pushed any further, 1 if
875  * it can.
876  */
877 int jbd2_journal_get_log_tail(journal_t *journal, tid_t *tid,
878                               unsigned long *block)
879 {
880         transaction_t *transaction;
881         int ret;
882
883         read_lock(&journal->j_state_lock);
884         spin_lock(&journal->j_list_lock);
885         transaction = journal->j_checkpoint_transactions;
886         if (transaction) {
887                 *tid = transaction->t_tid;
888                 *block = transaction->t_log_start;
889         } else if ((transaction = journal->j_committing_transaction) != NULL) {
890                 *tid = transaction->t_tid;
891                 *block = transaction->t_log_start;
892         } else if ((transaction = journal->j_running_transaction) != NULL) {
893                 *tid = transaction->t_tid;
894                 *block = journal->j_head;
895         } else {
896                 *tid = journal->j_transaction_sequence;
897                 *block = journal->j_head;
898         }
899         ret = tid_gt(*tid, journal->j_tail_sequence);
900         spin_unlock(&journal->j_list_lock);
901         read_unlock(&journal->j_state_lock);
902
903         return ret;
904 }
905
906 /*
907  * Update information in journal structure and in on disk journal superblock
908  * about log tail. This function does not check whether information passed in
909  * really pushes log tail further. It's responsibility of the caller to make
910  * sure provided log tail information is valid (e.g. by holding
911  * j_checkpoint_mutex all the time between computing log tail and calling this
912  * function as is the case with jbd2_cleanup_journal_tail()).
913  *
914  * Requires j_checkpoint_mutex
915  */
916 int __jbd2_update_log_tail(journal_t *journal, tid_t tid, unsigned long block)
917 {
918         unsigned long freed;
919         int ret;
920
921         BUG_ON(!mutex_is_locked(&journal->j_checkpoint_mutex));
922
923         /*
924          * We cannot afford for write to remain in drive's caches since as
925          * soon as we update j_tail, next transaction can start reusing journal
926          * space and if we lose sb update during power failure we'd replay
927          * old transaction with possibly newly overwritten data.
928          */
929         ret = jbd2_journal_update_sb_log_tail(journal, tid, block, WRITE_FUA);
930         if (ret)
931                 goto out;
932
933         write_lock(&journal->j_state_lock);
934         freed = block - journal->j_tail;
935         if (block < journal->j_tail)
936                 freed += journal->j_last - journal->j_first;
937
938         trace_jbd2_update_log_tail(journal, tid, block, freed);
939         jbd_debug(1,
940                   "Cleaning journal tail from %d to %d (offset %lu), "
941                   "freeing %lu\n",
942                   journal->j_tail_sequence, tid, block, freed);
943
944         journal->j_free += freed;
945         journal->j_tail_sequence = tid;
946         journal->j_tail = block;
947         write_unlock(&journal->j_state_lock);
948
949 out:
950         return ret;
951 }
952
953 /*
954  * This is a variation of __jbd2_update_log_tail which checks for validity of
955  * provided log tail and locks j_checkpoint_mutex. So it is safe against races
956  * with other threads updating log tail.
957  */
958 void jbd2_update_log_tail(journal_t *journal, tid_t tid, unsigned long block)
959 {
960         mutex_lock(&journal->j_checkpoint_mutex);
961         if (tid_gt(tid, journal->j_tail_sequence))
962                 __jbd2_update_log_tail(journal, tid, block);
963         mutex_unlock(&journal->j_checkpoint_mutex);
964 }
965
966 struct jbd2_stats_proc_session {
967         journal_t *journal;
968         struct transaction_stats_s *stats;
969         int start;
970         int max;
971 };
972
973 static void *jbd2_seq_info_start(struct seq_file *seq, loff_t *pos)
974 {
975         return *pos ? NULL : SEQ_START_TOKEN;
976 }
977
978 static void *jbd2_seq_info_next(struct seq_file *seq, void *v, loff_t *pos)
979 {
980         return NULL;
981 }
982
983 static int jbd2_seq_info_show(struct seq_file *seq, void *v)
984 {
985         struct jbd2_stats_proc_session *s = seq->private;
986
987         if (v != SEQ_START_TOKEN)
988                 return 0;
989         seq_printf(seq, "%lu transactions (%lu requested), "
990                    "each up to %u blocks\n",
991                    s->stats->ts_tid, s->stats->ts_requested,
992                    s->journal->j_max_transaction_buffers);
993         if (s->stats->ts_tid == 0)
994                 return 0;
995         seq_printf(seq, "average: \n  %ums waiting for transaction\n",
996             jiffies_to_msecs(s->stats->run.rs_wait / s->stats->ts_tid));
997         seq_printf(seq, "  %ums request delay\n",
998             (s->stats->ts_requested == 0) ? 0 :
999             jiffies_to_msecs(s->stats->run.rs_request_delay /
1000                              s->stats->ts_requested));
1001         seq_printf(seq, "  %ums running transaction\n",
1002             jiffies_to_msecs(s->stats->run.rs_running / s->stats->ts_tid));
1003         seq_printf(seq, "  %ums transaction was being locked\n",
1004             jiffies_to_msecs(s->stats->run.rs_locked / s->stats->ts_tid));
1005         seq_printf(seq, "  %ums flushing data (in ordered mode)\n",
1006             jiffies_to_msecs(s->stats->run.rs_flushing / s->stats->ts_tid));
1007         seq_printf(seq, "  %ums logging transaction\n",
1008             jiffies_to_msecs(s->stats->run.rs_logging / s->stats->ts_tid));
1009         seq_printf(seq, "  %lluus average transaction commit time\n",
1010                    div_u64(s->journal->j_average_commit_time, 1000));
1011         seq_printf(seq, "  %lu handles per transaction\n",
1012             s->stats->run.rs_handle_count / s->stats->ts_tid);
1013         seq_printf(seq, "  %lu blocks per transaction\n",
1014             s->stats->run.rs_blocks / s->stats->ts_tid);
1015         seq_printf(seq, "  %lu logged blocks per transaction\n",
1016             s->stats->run.rs_blocks_logged / s->stats->ts_tid);
1017         return 0;
1018 }
1019
1020 static void jbd2_seq_info_stop(struct seq_file *seq, void *v)
1021 {
1022 }
1023
1024 static const struct seq_operations jbd2_seq_info_ops = {
1025         .start  = jbd2_seq_info_start,
1026         .next   = jbd2_seq_info_next,
1027         .stop   = jbd2_seq_info_stop,
1028         .show   = jbd2_seq_info_show,
1029 };
1030
1031 static int jbd2_seq_info_open(struct inode *inode, struct file *file)
1032 {
1033         journal_t *journal = PDE_DATA(inode);
1034         struct jbd2_stats_proc_session *s;
1035         int rc, size;
1036
1037         s = kmalloc(sizeof(*s), GFP_KERNEL);
1038         if (s == NULL)
1039                 return -ENOMEM;
1040         size = sizeof(struct transaction_stats_s);
1041         s->stats = kmalloc(size, GFP_KERNEL);
1042         if (s->stats == NULL) {
1043                 kfree(s);
1044                 return -ENOMEM;
1045         }
1046         spin_lock(&journal->j_history_lock);
1047         memcpy(s->stats, &journal->j_stats, size);
1048         s->journal = journal;
1049         spin_unlock(&journal->j_history_lock);
1050
1051         rc = seq_open(file, &jbd2_seq_info_ops);
1052         if (rc == 0) {
1053                 struct seq_file *m = file->private_data;
1054                 m->private = s;
1055         } else {
1056                 kfree(s->stats);
1057                 kfree(s);
1058         }
1059         return rc;
1060
1061 }
1062
1063 static int jbd2_seq_info_release(struct inode *inode, struct file *file)
1064 {
1065         struct seq_file *seq = file->private_data;
1066         struct jbd2_stats_proc_session *s = seq->private;
1067         kfree(s->stats);
1068         kfree(s);
1069         return seq_release(inode, file);
1070 }
1071
1072 static const struct file_operations jbd2_seq_info_fops = {
1073         .owner          = THIS_MODULE,
1074         .open           = jbd2_seq_info_open,
1075         .read           = seq_read,
1076         .llseek         = seq_lseek,
1077         .release        = jbd2_seq_info_release,
1078 };
1079
1080 static struct proc_dir_entry *proc_jbd2_stats;
1081
1082 static void jbd2_stats_proc_init(journal_t *journal)
1083 {
1084         journal->j_proc_entry = proc_mkdir(journal->j_devname, proc_jbd2_stats);
1085         if (journal->j_proc_entry) {
1086                 proc_create_data("info", S_IRUGO, journal->j_proc_entry,
1087                                  &jbd2_seq_info_fops, journal);
1088         }
1089 }
1090
1091 static void jbd2_stats_proc_exit(journal_t *journal)
1092 {
1093         remove_proc_entry("info", journal->j_proc_entry);
1094         remove_proc_entry(journal->j_devname, proc_jbd2_stats);
1095 }
1096
1097 /*
1098  * Management for journal control blocks: functions to create and
1099  * destroy journal_t structures, and to initialise and read existing
1100  * journal blocks from disk.  */
1101
1102 /* First: create and setup a journal_t object in memory.  We initialise
1103  * very few fields yet: that has to wait until we have created the
1104  * journal structures from from scratch, or loaded them from disk. */
1105
1106 static journal_t *journal_init_common(struct block_device *bdev,
1107                         struct block_device *fs_dev,
1108                         unsigned long long start, int len, int blocksize)
1109 {
1110         static struct lock_class_key jbd2_trans_commit_key;
1111         journal_t *journal;
1112         int err;
1113         struct buffer_head *bh;
1114         int n;
1115
1116         journal = kzalloc(sizeof(*journal), GFP_KERNEL);
1117         if (!journal)
1118                 return NULL;
1119
1120         init_waitqueue_head(&journal->j_wait_transaction_locked);
1121         init_waitqueue_head(&journal->j_wait_done_commit);
1122         init_waitqueue_head(&journal->j_wait_commit);
1123         init_waitqueue_head(&journal->j_wait_updates);
1124         init_waitqueue_head(&journal->j_wait_reserved);
1125         mutex_init(&journal->j_barrier);
1126         mutex_init(&journal->j_checkpoint_mutex);
1127         spin_lock_init(&journal->j_revoke_lock);
1128         spin_lock_init(&journal->j_list_lock);
1129         rwlock_init(&journal->j_state_lock);
1130
1131         journal->j_commit_interval = (HZ * JBD2_DEFAULT_MAX_COMMIT_AGE);
1132         journal->j_min_batch_time = 0;
1133         journal->j_max_batch_time = 15000; /* 15ms */
1134         atomic_set(&journal->j_reserved_credits, 0);
1135
1136         /* The journal is marked for error until we succeed with recovery! */
1137         journal->j_flags = JBD2_ABORT;
1138
1139         /* Set up a default-sized revoke table for the new mount. */
1140         err = jbd2_journal_init_revoke(journal, JOURNAL_REVOKE_DEFAULT_HASH);
1141         if (err)
1142                 goto err_cleanup;
1143
1144         spin_lock_init(&journal->j_history_lock);
1145
1146         lockdep_init_map(&journal->j_trans_commit_map, "jbd2_handle",
1147                          &jbd2_trans_commit_key, 0);
1148
1149         /* journal descriptor can store up to n blocks -bzzz */
1150         journal->j_blocksize = blocksize;
1151         journal->j_dev = bdev;
1152         journal->j_fs_dev = fs_dev;
1153         journal->j_blk_offset = start;
1154         journal->j_maxlen = len;
1155         n = journal->j_blocksize / sizeof(journal_block_tag_t);
1156         journal->j_wbufsize = n;
1157         journal->j_wbuf = kmalloc_array(n, sizeof(struct buffer_head *),
1158                                         GFP_KERNEL);
1159         if (!journal->j_wbuf)
1160                 goto err_cleanup;
1161
1162         bh = getblk_unmovable(journal->j_dev, start, journal->j_blocksize);
1163         if (!bh) {
1164                 pr_err("%s: Cannot get buffer for journal superblock\n",
1165                         __func__);
1166                 goto err_cleanup;
1167         }
1168         journal->j_sb_buffer = bh;
1169         journal->j_superblock = (journal_superblock_t *)bh->b_data;
1170
1171         return journal;
1172
1173 err_cleanup:
1174         kfree(journal->j_wbuf);
1175         jbd2_journal_destroy_revoke(journal);
1176         kfree(journal);
1177         return NULL;
1178 }
1179
1180 /* jbd2_journal_init_dev and jbd2_journal_init_inode:
1181  *
1182  * Create a journal structure assigned some fixed set of disk blocks to
1183  * the journal.  We don't actually touch those disk blocks yet, but we
1184  * need to set up all of the mapping information to tell the journaling
1185  * system where the journal blocks are.
1186  *
1187  */
1188
1189 /**
1190  *  journal_t * jbd2_journal_init_dev() - creates and initialises a journal structure
1191  *  @bdev: Block device on which to create the journal
1192  *  @fs_dev: Device which hold journalled filesystem for this journal.
1193  *  @start: Block nr Start of journal.
1194  *  @len:  Length of the journal in blocks.
1195  *  @blocksize: blocksize of journalling device
1196  *
1197  *  Returns: a newly created journal_t *
1198  *
1199  *  jbd2_journal_init_dev creates a journal which maps a fixed contiguous
1200  *  range of blocks on an arbitrary block device.
1201  *
1202  */
1203 journal_t *jbd2_journal_init_dev(struct block_device *bdev,
1204                         struct block_device *fs_dev,
1205                         unsigned long long start, int len, int blocksize)
1206 {
1207         journal_t *journal;
1208
1209         journal = journal_init_common(bdev, fs_dev, start, len, blocksize);
1210         if (!journal)
1211                 return NULL;
1212
1213         bdevname(journal->j_dev, journal->j_devname);
1214         strreplace(journal->j_devname, '/', '!');
1215         jbd2_stats_proc_init(journal);
1216
1217         return journal;
1218 }
1219
1220 /**
1221  *  journal_t * jbd2_journal_init_inode () - creates a journal which maps to a inode.
1222  *  @inode: An inode to create the journal in
1223  *
1224  * jbd2_journal_init_inode creates a journal which maps an on-disk inode as
1225  * the journal.  The inode must exist already, must support bmap() and
1226  * must have all data blocks preallocated.
1227  */
1228 journal_t *jbd2_journal_init_inode(struct inode *inode)
1229 {
1230         journal_t *journal;
1231         char *p;
1232         unsigned long long blocknr;
1233
1234         blocknr = bmap(inode, 0);
1235         if (!blocknr) {
1236                 pr_err("%s: Cannot locate journal superblock\n",
1237                         __func__);
1238                 return NULL;
1239         }
1240
1241         jbd_debug(1, "JBD2: inode %s/%ld, size %lld, bits %d, blksize %ld\n",
1242                   inode->i_sb->s_id, inode->i_ino, (long long) inode->i_size,
1243                   inode->i_sb->s_blocksize_bits, inode->i_sb->s_blocksize);
1244
1245         journal = journal_init_common(inode->i_sb->s_bdev, inode->i_sb->s_bdev,
1246                         blocknr, inode->i_size >> inode->i_sb->s_blocksize_bits,
1247                         inode->i_sb->s_blocksize);
1248         if (!journal)
1249                 return NULL;
1250
1251         journal->j_inode = inode;
1252         bdevname(journal->j_dev, journal->j_devname);
1253         p = strreplace(journal->j_devname, '/', '!');
1254         sprintf(p, "-%lu", journal->j_inode->i_ino);
1255         jbd2_stats_proc_init(journal);
1256
1257         return journal;
1258 }
1259
1260 /*
1261  * If the journal init or create aborts, we need to mark the journal
1262  * superblock as being NULL to prevent the journal destroy from writing
1263  * back a bogus superblock.
1264  */
1265 static void journal_fail_superblock (journal_t *journal)
1266 {
1267         struct buffer_head *bh = journal->j_sb_buffer;
1268         brelse(bh);
1269         journal->j_sb_buffer = NULL;
1270 }
1271
1272 /*
1273  * Given a journal_t structure, initialise the various fields for
1274  * startup of a new journaling session.  We use this both when creating
1275  * a journal, and after recovering an old journal to reset it for
1276  * subsequent use.
1277  */
1278
1279 static int journal_reset(journal_t *journal)
1280 {
1281         journal_superblock_t *sb = journal->j_superblock;
1282         unsigned long long first, last;
1283
1284         first = be32_to_cpu(sb->s_first);
1285         last = be32_to_cpu(sb->s_maxlen);
1286         if (first + JBD2_MIN_JOURNAL_BLOCKS > last + 1) {
1287                 printk(KERN_ERR "JBD2: Journal too short (blocks %llu-%llu).\n",
1288                        first, last);
1289                 journal_fail_superblock(journal);
1290                 return -EINVAL;
1291         }
1292
1293         journal->j_first = first;
1294         journal->j_last = last;
1295
1296         journal->j_head = first;
1297         journal->j_tail = first;
1298         journal->j_free = last - first;
1299
1300         journal->j_tail_sequence = journal->j_transaction_sequence;
1301         journal->j_commit_sequence = journal->j_transaction_sequence - 1;
1302         journal->j_commit_request = journal->j_commit_sequence;
1303
1304         journal->j_max_transaction_buffers = journal->j_maxlen / 4;
1305
1306         /*
1307          * As a special case, if the on-disk copy is already marked as needing
1308          * no recovery (s_start == 0), then we can safely defer the superblock
1309          * update until the next commit by setting JBD2_FLUSHED.  This avoids
1310          * attempting a write to a potential-readonly device.
1311          */
1312         if (sb->s_start == 0) {
1313                 jbd_debug(1, "JBD2: Skipping superblock update on recovered sb "
1314                         "(start %ld, seq %d, errno %d)\n",
1315                         journal->j_tail, journal->j_tail_sequence,
1316                         journal->j_errno);
1317                 journal->j_flags |= JBD2_FLUSHED;
1318         } else {
1319                 /* Lock here to make assertions happy... */
1320                 mutex_lock(&journal->j_checkpoint_mutex);
1321                 /*
1322                  * Update log tail information. We use WRITE_FUA since new
1323                  * transaction will start reusing journal space and so we
1324                  * must make sure information about current log tail is on
1325                  * disk before that.
1326                  */
1327                 jbd2_journal_update_sb_log_tail(journal,
1328                                                 journal->j_tail_sequence,
1329                                                 journal->j_tail,
1330                                                 WRITE_FUA);
1331                 mutex_unlock(&journal->j_checkpoint_mutex);
1332         }
1333         return jbd2_journal_start_thread(journal);
1334 }
1335
1336 static int jbd2_write_superblock(journal_t *journal, int write_flags)
1337 {
1338         struct buffer_head *bh = journal->j_sb_buffer;
1339         journal_superblock_t *sb = journal->j_superblock;
1340         int ret;
1341
1342         /* Buffer got discarded which means block device got invalidated */
1343         if (!buffer_mapped(bh))
1344                 return -EIO;
1345
1346         trace_jbd2_write_superblock(journal, write_flags);
1347         if (!(journal->j_flags & JBD2_BARRIER))
1348                 write_flags &= ~(REQ_FUA | REQ_PREFLUSH);
1349         lock_buffer(bh);
1350         if (buffer_write_io_error(bh)) {
1351                 /*
1352                  * Oh, dear.  A previous attempt to write the journal
1353                  * superblock failed.  This could happen because the
1354                  * USB device was yanked out.  Or it could happen to
1355                  * be a transient write error and maybe the block will
1356                  * be remapped.  Nothing we can do but to retry the
1357                  * write and hope for the best.
1358                  */
1359                 printk(KERN_ERR "JBD2: previous I/O error detected "
1360                        "for journal superblock update for %s.\n",
1361                        journal->j_devname);
1362                 clear_buffer_write_io_error(bh);
1363                 set_buffer_uptodate(bh);
1364         }
1365         jbd2_superblock_csum_set(journal, sb);
1366         get_bh(bh);
1367         bh->b_end_io = end_buffer_write_sync;
1368         ret = submit_bh(REQ_OP_WRITE, write_flags, bh);
1369         wait_on_buffer(bh);
1370         if (buffer_write_io_error(bh)) {
1371                 clear_buffer_write_io_error(bh);
1372                 set_buffer_uptodate(bh);
1373                 ret = -EIO;
1374         }
1375         if (ret) {
1376                 printk(KERN_ERR "JBD2: Error %d detected when updating "
1377                        "journal superblock for %s.\n", ret,
1378                        journal->j_devname);
1379                 jbd2_journal_abort(journal, ret);
1380         }
1381
1382         return ret;
1383 }
1384
1385 /**
1386  * jbd2_journal_update_sb_log_tail() - Update log tail in journal sb on disk.
1387  * @journal: The journal to update.
1388  * @tail_tid: TID of the new transaction at the tail of the log
1389  * @tail_block: The first block of the transaction at the tail of the log
1390  * @write_op: With which operation should we write the journal sb
1391  *
1392  * Update a journal's superblock information about log tail and write it to
1393  * disk, waiting for the IO to complete.
1394  */
1395 int jbd2_journal_update_sb_log_tail(journal_t *journal, tid_t tail_tid,
1396                                      unsigned long tail_block, int write_op)
1397 {
1398         journal_superblock_t *sb = journal->j_superblock;
1399         int ret;
1400
1401         if (is_journal_aborted(journal))
1402                 return -EIO;
1403
1404         BUG_ON(!mutex_is_locked(&journal->j_checkpoint_mutex));
1405         jbd_debug(1, "JBD2: updating superblock (start %lu, seq %u)\n",
1406                   tail_block, tail_tid);
1407
1408         sb->s_sequence = cpu_to_be32(tail_tid);
1409         sb->s_start    = cpu_to_be32(tail_block);
1410
1411         ret = jbd2_write_superblock(journal, write_op);
1412         if (ret)
1413                 goto out;
1414
1415         /* Log is no longer empty */
1416         write_lock(&journal->j_state_lock);
1417         WARN_ON(!sb->s_sequence);
1418         journal->j_flags &= ~JBD2_FLUSHED;
1419         write_unlock(&journal->j_state_lock);
1420
1421 out:
1422         return ret;
1423 }
1424
1425 /**
1426  * jbd2_mark_journal_empty() - Mark on disk journal as empty.
1427  * @journal: The journal to update.
1428  * @write_op: With which operation should we write the journal sb
1429  *
1430  * Update a journal's dynamic superblock fields to show that journal is empty.
1431  * Write updated superblock to disk waiting for IO to complete.
1432  */
1433 static void jbd2_mark_journal_empty(journal_t *journal, int write_op)
1434 {
1435         journal_superblock_t *sb = journal->j_superblock;
1436
1437         BUG_ON(!mutex_is_locked(&journal->j_checkpoint_mutex));
1438         read_lock(&journal->j_state_lock);
1439         /* Is it already empty? */
1440         if (sb->s_start == 0) {
1441                 read_unlock(&journal->j_state_lock);
1442                 return;
1443         }
1444         jbd_debug(1, "JBD2: Marking journal as empty (seq %d)\n",
1445                   journal->j_tail_sequence);
1446
1447         sb->s_sequence = cpu_to_be32(journal->j_tail_sequence);
1448         sb->s_start    = cpu_to_be32(0);
1449         read_unlock(&journal->j_state_lock);
1450
1451         jbd2_write_superblock(journal, write_op);
1452
1453         /* Log is no longer empty */
1454         write_lock(&journal->j_state_lock);
1455         journal->j_flags |= JBD2_FLUSHED;
1456         write_unlock(&journal->j_state_lock);
1457 }
1458
1459
1460 /**
1461  * jbd2_journal_update_sb_errno() - Update error in the journal.
1462  * @journal: The journal to update.
1463  *
1464  * Update a journal's errno.  Write updated superblock to disk waiting for IO
1465  * to complete.
1466  */
1467 void jbd2_journal_update_sb_errno(journal_t *journal)
1468 {
1469         journal_superblock_t *sb = journal->j_superblock;
1470
1471         read_lock(&journal->j_state_lock);
1472         jbd_debug(1, "JBD2: updating superblock error (errno %d)\n",
1473                   journal->j_errno);
1474         sb->s_errno    = cpu_to_be32(journal->j_errno);
1475         read_unlock(&journal->j_state_lock);
1476
1477         jbd2_write_superblock(journal, WRITE_FUA);
1478 }
1479 EXPORT_SYMBOL(jbd2_journal_update_sb_errno);
1480
1481 /*
1482  * Read the superblock for a given journal, performing initial
1483  * validation of the format.
1484  */
1485 static int journal_get_superblock(journal_t *journal)
1486 {
1487         struct buffer_head *bh;
1488         journal_superblock_t *sb;
1489         int err = -EIO;
1490
1491         bh = journal->j_sb_buffer;
1492
1493         J_ASSERT(bh != NULL);
1494         if (!buffer_uptodate(bh)) {
1495                 ll_rw_block(REQ_OP_READ, 0, 1, &bh);
1496                 wait_on_buffer(bh);
1497                 if (!buffer_uptodate(bh)) {
1498                         printk(KERN_ERR
1499                                 "JBD2: IO error reading journal superblock\n");
1500                         goto out;
1501                 }
1502         }
1503
1504         if (buffer_verified(bh))
1505                 return 0;
1506
1507         sb = journal->j_superblock;
1508
1509         err = -EINVAL;
1510
1511         if (sb->s_header.h_magic != cpu_to_be32(JBD2_MAGIC_NUMBER) ||
1512             sb->s_blocksize != cpu_to_be32(journal->j_blocksize)) {
1513                 printk(KERN_WARNING "JBD2: no valid journal superblock found\n");
1514                 goto out;
1515         }
1516
1517         switch(be32_to_cpu(sb->s_header.h_blocktype)) {
1518         case JBD2_SUPERBLOCK_V1:
1519                 journal->j_format_version = 1;
1520                 break;
1521         case JBD2_SUPERBLOCK_V2:
1522                 journal->j_format_version = 2;
1523                 break;
1524         default:
1525                 printk(KERN_WARNING "JBD2: unrecognised superblock format ID\n");
1526                 goto out;
1527         }
1528
1529         if (be32_to_cpu(sb->s_maxlen) < journal->j_maxlen)
1530                 journal->j_maxlen = be32_to_cpu(sb->s_maxlen);
1531         else if (be32_to_cpu(sb->s_maxlen) > journal->j_maxlen) {
1532                 printk(KERN_WARNING "JBD2: journal file too short\n");
1533                 goto out;
1534         }
1535
1536         if (be32_to_cpu(sb->s_first) == 0 ||
1537             be32_to_cpu(sb->s_first) >= journal->j_maxlen) {
1538                 printk(KERN_WARNING
1539                         "JBD2: Invalid start block of journal: %u\n",
1540                         be32_to_cpu(sb->s_first));
1541                 goto out;
1542         }
1543
1544         if (jbd2_has_feature_csum2(journal) &&
1545             jbd2_has_feature_csum3(journal)) {
1546                 /* Can't have checksum v2 and v3 at the same time! */
1547                 printk(KERN_ERR "JBD2: Can't enable checksumming v2 and v3 "
1548                        "at the same time!\n");
1549                 goto out;
1550         }
1551
1552         if (jbd2_journal_has_csum_v2or3_feature(journal) &&
1553             jbd2_has_feature_checksum(journal)) {
1554                 /* Can't have checksum v1 and v2 on at the same time! */
1555                 printk(KERN_ERR "JBD2: Can't enable checksumming v1 and v2/3 "
1556                        "at the same time!\n");
1557                 goto out;
1558         }
1559
1560         if (!jbd2_verify_csum_type(journal, sb)) {
1561                 printk(KERN_ERR "JBD2: Unknown checksum type\n");
1562                 goto out;
1563         }
1564
1565         /* Load the checksum driver */
1566         if (jbd2_journal_has_csum_v2or3_feature(journal)) {
1567                 journal->j_chksum_driver = crypto_alloc_shash("crc32c", 0, 0);
1568                 if (IS_ERR(journal->j_chksum_driver)) {
1569                         printk(KERN_ERR "JBD2: Cannot load crc32c driver.\n");
1570                         err = PTR_ERR(journal->j_chksum_driver);
1571                         journal->j_chksum_driver = NULL;
1572                         goto out;
1573                 }
1574         }
1575
1576         /* Check superblock checksum */
1577         if (!jbd2_superblock_csum_verify(journal, sb)) {
1578                 printk(KERN_ERR "JBD2: journal checksum error\n");
1579                 err = -EFSBADCRC;
1580                 goto out;
1581         }
1582
1583         /* Precompute checksum seed for all metadata */
1584         if (jbd2_journal_has_csum_v2or3(journal))
1585                 journal->j_csum_seed = jbd2_chksum(journal, ~0, sb->s_uuid,
1586                                                    sizeof(sb->s_uuid));
1587
1588         set_buffer_verified(bh);
1589
1590         return 0;
1591
1592 out:
1593         journal_fail_superblock(journal);
1594         return err;
1595 }
1596
1597 /*
1598  * Load the on-disk journal superblock and read the key fields into the
1599  * journal_t.
1600  */
1601
1602 static int load_superblock(journal_t *journal)
1603 {
1604         int err;
1605         journal_superblock_t *sb;
1606
1607         err = journal_get_superblock(journal);
1608         if (err)
1609                 return err;
1610
1611         sb = journal->j_superblock;
1612
1613         journal->j_tail_sequence = be32_to_cpu(sb->s_sequence);
1614         journal->j_tail = be32_to_cpu(sb->s_start);
1615         journal->j_first = be32_to_cpu(sb->s_first);
1616         journal->j_last = be32_to_cpu(sb->s_maxlen);
1617         journal->j_errno = be32_to_cpu(sb->s_errno);
1618
1619         return 0;
1620 }
1621
1622
1623 /**
1624  * int jbd2_journal_load() - Read journal from disk.
1625  * @journal: Journal to act on.
1626  *
1627  * Given a journal_t structure which tells us which disk blocks contain
1628  * a journal, read the journal from disk to initialise the in-memory
1629  * structures.
1630  */
1631 int jbd2_journal_load(journal_t *journal)
1632 {
1633         int err;
1634         journal_superblock_t *sb;
1635
1636         err = load_superblock(journal);
1637         if (err)
1638                 return err;
1639
1640         sb = journal->j_superblock;
1641         /* If this is a V2 superblock, then we have to check the
1642          * features flags on it. */
1643
1644         if (journal->j_format_version >= 2) {
1645                 if ((sb->s_feature_ro_compat &
1646                      ~cpu_to_be32(JBD2_KNOWN_ROCOMPAT_FEATURES)) ||
1647                     (sb->s_feature_incompat &
1648                      ~cpu_to_be32(JBD2_KNOWN_INCOMPAT_FEATURES))) {
1649                         printk(KERN_WARNING
1650                                 "JBD2: Unrecognised features on journal\n");
1651                         return -EINVAL;
1652                 }
1653         }
1654
1655         /*
1656          * Create a slab for this blocksize
1657          */
1658         err = jbd2_journal_create_slab(be32_to_cpu(sb->s_blocksize));
1659         if (err)
1660                 return err;
1661
1662         /* Let the recovery code check whether it needs to recover any
1663          * data from the journal. */
1664         if (jbd2_journal_recover(journal))
1665                 goto recovery_error;
1666
1667         if (journal->j_failed_commit) {
1668                 printk(KERN_ERR "JBD2: journal transaction %u on %s "
1669                        "is corrupt.\n", journal->j_failed_commit,
1670                        journal->j_devname);
1671                 return -EFSCORRUPTED;
1672         }
1673         /*
1674          * clear JBD2_ABORT flag initialized in journal_init_common
1675          * here to update log tail information with the newest seq.
1676          */
1677         journal->j_flags &= ~JBD2_ABORT;
1678
1679         /* OK, we've finished with the dynamic journal bits:
1680          * reinitialise the dynamic contents of the superblock in memory
1681          * and reset them on disk. */
1682         if (journal_reset(journal))
1683                 goto recovery_error;
1684
1685         journal->j_flags |= JBD2_LOADED;
1686         return 0;
1687
1688 recovery_error:
1689         printk(KERN_WARNING "JBD2: recovery failed\n");
1690         return -EIO;
1691 }
1692
1693 /**
1694  * void jbd2_journal_destroy() - Release a journal_t structure.
1695  * @journal: Journal to act on.
1696  *
1697  * Release a journal_t structure once it is no longer in use by the
1698  * journaled object.
1699  * Return <0 if we couldn't clean up the journal.
1700  */
1701 int jbd2_journal_destroy(journal_t *journal)
1702 {
1703         int err = 0;
1704
1705         /* Wait for the commit thread to wake up and die. */
1706         journal_kill_thread(journal);
1707
1708         /* Force a final log commit */
1709         if (journal->j_running_transaction)
1710                 jbd2_journal_commit_transaction(journal);
1711
1712         /* Force any old transactions to disk */
1713
1714         /* Totally anal locking here... */
1715         spin_lock(&journal->j_list_lock);
1716         while (journal->j_checkpoint_transactions != NULL) {
1717                 spin_unlock(&journal->j_list_lock);
1718                 mutex_lock(&journal->j_checkpoint_mutex);
1719                 err = jbd2_log_do_checkpoint(journal);
1720                 mutex_unlock(&journal->j_checkpoint_mutex);
1721                 /*
1722                  * If checkpointing failed, just free the buffers to avoid
1723                  * looping forever
1724                  */
1725                 if (err) {
1726                         jbd2_journal_destroy_checkpoint(journal);
1727                         spin_lock(&journal->j_list_lock);
1728                         break;
1729                 }
1730                 spin_lock(&journal->j_list_lock);
1731         }
1732
1733         J_ASSERT(journal->j_running_transaction == NULL);
1734         J_ASSERT(journal->j_committing_transaction == NULL);
1735         J_ASSERT(journal->j_checkpoint_transactions == NULL);
1736         spin_unlock(&journal->j_list_lock);
1737
1738         if (journal->j_sb_buffer) {
1739                 if (!is_journal_aborted(journal)) {
1740                         mutex_lock(&journal->j_checkpoint_mutex);
1741
1742                         write_lock(&journal->j_state_lock);
1743                         journal->j_tail_sequence =
1744                                 ++journal->j_transaction_sequence;
1745                         write_unlock(&journal->j_state_lock);
1746
1747                         jbd2_mark_journal_empty(journal, WRITE_FLUSH_FUA);
1748                         mutex_unlock(&journal->j_checkpoint_mutex);
1749                 } else
1750                         err = -EIO;
1751                 brelse(journal->j_sb_buffer);
1752         }
1753
1754         if (journal->j_proc_entry)
1755                 jbd2_stats_proc_exit(journal);
1756         iput(journal->j_inode);
1757         if (journal->j_revoke)
1758                 jbd2_journal_destroy_revoke(journal);
1759         if (journal->j_chksum_driver)
1760                 crypto_free_shash(journal->j_chksum_driver);
1761         kfree(journal->j_wbuf);
1762         kfree(journal);
1763
1764         return err;
1765 }
1766
1767
1768 /**
1769  *int jbd2_journal_check_used_features () - Check if features specified are used.
1770  * @journal: Journal to check.
1771  * @compat: bitmask of compatible features
1772  * @ro: bitmask of features that force read-only mount
1773  * @incompat: bitmask of incompatible features
1774  *
1775  * Check whether the journal uses all of a given set of
1776  * features.  Return true (non-zero) if it does.
1777  **/
1778
1779 int jbd2_journal_check_used_features (journal_t *journal, unsigned long compat,
1780                                  unsigned long ro, unsigned long incompat)
1781 {
1782         journal_superblock_t *sb;
1783
1784         if (!compat && !ro && !incompat)
1785                 return 1;
1786         /* Load journal superblock if it is not loaded yet. */
1787         if (journal->j_format_version == 0 &&
1788             journal_get_superblock(journal) != 0)
1789                 return 0;
1790         if (journal->j_format_version == 1)
1791                 return 0;
1792
1793         sb = journal->j_superblock;
1794
1795         if (((be32_to_cpu(sb->s_feature_compat) & compat) == compat) &&
1796             ((be32_to_cpu(sb->s_feature_ro_compat) & ro) == ro) &&
1797             ((be32_to_cpu(sb->s_feature_incompat) & incompat) == incompat))
1798                 return 1;
1799
1800         return 0;
1801 }
1802
1803 /**
1804  * int jbd2_journal_check_available_features() - Check feature set in journalling layer
1805  * @journal: Journal to check.
1806  * @compat: bitmask of compatible features
1807  * @ro: bitmask of features that force read-only mount
1808  * @incompat: bitmask of incompatible features
1809  *
1810  * Check whether the journaling code supports the use of
1811  * all of a given set of features on this journal.  Return true
1812  * (non-zero) if it can. */
1813
1814 int jbd2_journal_check_available_features (journal_t *journal, unsigned long compat,
1815                                       unsigned long ro, unsigned long incompat)
1816 {
1817         if (!compat && !ro && !incompat)
1818                 return 1;
1819
1820         /* We can support any known requested features iff the
1821          * superblock is in version 2.  Otherwise we fail to support any
1822          * extended sb features. */
1823
1824         if (journal->j_format_version != 2)
1825                 return 0;
1826
1827         if ((compat   & JBD2_KNOWN_COMPAT_FEATURES) == compat &&
1828             (ro       & JBD2_KNOWN_ROCOMPAT_FEATURES) == ro &&
1829             (incompat & JBD2_KNOWN_INCOMPAT_FEATURES) == incompat)
1830                 return 1;
1831
1832         return 0;
1833 }
1834
1835 /**
1836  * int jbd2_journal_set_features () - Mark a given journal feature in the superblock
1837  * @journal: Journal to act on.
1838  * @compat: bitmask of compatible features
1839  * @ro: bitmask of features that force read-only mount
1840  * @incompat: bitmask of incompatible features
1841  *
1842  * Mark a given journal feature as present on the
1843  * superblock.  Returns true if the requested features could be set.
1844  *
1845  */
1846
1847 int jbd2_journal_set_features (journal_t *journal, unsigned long compat,
1848                           unsigned long ro, unsigned long incompat)
1849 {
1850 #define INCOMPAT_FEATURE_ON(f) \
1851                 ((incompat & (f)) && !(sb->s_feature_incompat & cpu_to_be32(f)))
1852 #define COMPAT_FEATURE_ON(f) \
1853                 ((compat & (f)) && !(sb->s_feature_compat & cpu_to_be32(f)))
1854         journal_superblock_t *sb;
1855
1856         if (jbd2_journal_check_used_features(journal, compat, ro, incompat))
1857                 return 1;
1858
1859         if (!jbd2_journal_check_available_features(journal, compat, ro, incompat))
1860                 return 0;
1861
1862         /* If enabling v2 checksums, turn on v3 instead */
1863         if (incompat & JBD2_FEATURE_INCOMPAT_CSUM_V2) {
1864                 incompat &= ~JBD2_FEATURE_INCOMPAT_CSUM_V2;
1865                 incompat |= JBD2_FEATURE_INCOMPAT_CSUM_V3;
1866         }
1867
1868         /* Asking for checksumming v3 and v1?  Only give them v3. */
1869         if (incompat & JBD2_FEATURE_INCOMPAT_CSUM_V3 &&
1870             compat & JBD2_FEATURE_COMPAT_CHECKSUM)
1871                 compat &= ~JBD2_FEATURE_COMPAT_CHECKSUM;
1872
1873         jbd_debug(1, "Setting new features 0x%lx/0x%lx/0x%lx\n",
1874                   compat, ro, incompat);
1875
1876         sb = journal->j_superblock;
1877
1878         /* If enabling v3 checksums, update superblock */
1879         if (INCOMPAT_FEATURE_ON(JBD2_FEATURE_INCOMPAT_CSUM_V3)) {
1880                 sb->s_checksum_type = JBD2_CRC32C_CHKSUM;
1881                 sb->s_feature_compat &=
1882                         ~cpu_to_be32(JBD2_FEATURE_COMPAT_CHECKSUM);
1883
1884                 /* Load the checksum driver */
1885                 if (journal->j_chksum_driver == NULL) {
1886                         journal->j_chksum_driver = crypto_alloc_shash("crc32c",
1887                                                                       0, 0);
1888                         if (IS_ERR(journal->j_chksum_driver)) {
1889                                 printk(KERN_ERR "JBD2: Cannot load crc32c "
1890                                        "driver.\n");
1891                                 journal->j_chksum_driver = NULL;
1892                                 return 0;
1893                         }
1894
1895                         /* Precompute checksum seed for all metadata */
1896                         journal->j_csum_seed = jbd2_chksum(journal, ~0,
1897                                                            sb->s_uuid,
1898                                                            sizeof(sb->s_uuid));
1899                 }
1900         }
1901
1902         /* If enabling v1 checksums, downgrade superblock */
1903         if (COMPAT_FEATURE_ON(JBD2_FEATURE_COMPAT_CHECKSUM))
1904                 sb->s_feature_incompat &=
1905                         ~cpu_to_be32(JBD2_FEATURE_INCOMPAT_CSUM_V2 |
1906                                      JBD2_FEATURE_INCOMPAT_CSUM_V3);
1907
1908         sb->s_feature_compat    |= cpu_to_be32(compat);
1909         sb->s_feature_ro_compat |= cpu_to_be32(ro);
1910         sb->s_feature_incompat  |= cpu_to_be32(incompat);
1911
1912         return 1;
1913 #undef COMPAT_FEATURE_ON
1914 #undef INCOMPAT_FEATURE_ON
1915 }
1916
1917 /*
1918  * jbd2_journal_clear_features () - Clear a given journal feature in the
1919  *                                  superblock
1920  * @journal: Journal to act on.
1921  * @compat: bitmask of compatible features
1922  * @ro: bitmask of features that force read-only mount
1923  * @incompat: bitmask of incompatible features
1924  *
1925  * Clear a given journal feature as present on the
1926  * superblock.
1927  */
1928 void jbd2_journal_clear_features(journal_t *journal, unsigned long compat,
1929                                 unsigned long ro, unsigned long incompat)
1930 {
1931         journal_superblock_t *sb;
1932
1933         jbd_debug(1, "Clear features 0x%lx/0x%lx/0x%lx\n",
1934                   compat, ro, incompat);
1935
1936         sb = journal->j_superblock;
1937
1938         sb->s_feature_compat    &= ~cpu_to_be32(compat);
1939         sb->s_feature_ro_compat &= ~cpu_to_be32(ro);
1940         sb->s_feature_incompat  &= ~cpu_to_be32(incompat);
1941 }
1942 EXPORT_SYMBOL(jbd2_journal_clear_features);
1943
1944 /**
1945  * int jbd2_journal_flush () - Flush journal
1946  * @journal: Journal to act on.
1947  *
1948  * Flush all data for a given journal to disk and empty the journal.
1949  * Filesystems can use this when remounting readonly to ensure that
1950  * recovery does not need to happen on remount.
1951  */
1952
1953 int jbd2_journal_flush(journal_t *journal)
1954 {
1955         int err = 0;
1956         transaction_t *transaction = NULL;
1957
1958         write_lock(&journal->j_state_lock);
1959
1960         /* Force everything buffered to the log... */
1961         if (journal->j_running_transaction) {
1962                 transaction = journal->j_running_transaction;
1963                 __jbd2_log_start_commit(journal, transaction->t_tid);
1964         } else if (journal->j_committing_transaction)
1965                 transaction = journal->j_committing_transaction;
1966
1967         /* Wait for the log commit to complete... */
1968         if (transaction) {
1969                 tid_t tid = transaction->t_tid;
1970
1971                 write_unlock(&journal->j_state_lock);
1972                 jbd2_log_wait_commit(journal, tid);
1973         } else {
1974                 write_unlock(&journal->j_state_lock);
1975         }
1976
1977         /* ...and flush everything in the log out to disk. */
1978         spin_lock(&journal->j_list_lock);
1979         while (!err && journal->j_checkpoint_transactions != NULL) {
1980                 spin_unlock(&journal->j_list_lock);
1981                 mutex_lock(&journal->j_checkpoint_mutex);
1982                 err = jbd2_log_do_checkpoint(journal);
1983                 mutex_unlock(&journal->j_checkpoint_mutex);
1984                 spin_lock(&journal->j_list_lock);
1985         }
1986         spin_unlock(&journal->j_list_lock);
1987
1988         if (is_journal_aborted(journal))
1989                 return -EIO;
1990
1991         mutex_lock(&journal->j_checkpoint_mutex);
1992         if (!err) {
1993                 err = jbd2_cleanup_journal_tail(journal);
1994                 if (err < 0) {
1995                         mutex_unlock(&journal->j_checkpoint_mutex);
1996                         goto out;
1997                 }
1998                 err = 0;
1999         }
2000
2001         /* Finally, mark the journal as really needing no recovery.
2002          * This sets s_start==0 in the underlying superblock, which is
2003          * the magic code for a fully-recovered superblock.  Any future
2004          * commits of data to the journal will restore the current
2005          * s_start value. */
2006         jbd2_mark_journal_empty(journal, WRITE_FUA);
2007         mutex_unlock(&journal->j_checkpoint_mutex);
2008         write_lock(&journal->j_state_lock);
2009         J_ASSERT(!journal->j_running_transaction);
2010         J_ASSERT(!journal->j_committing_transaction);
2011         J_ASSERT(!journal->j_checkpoint_transactions);
2012         J_ASSERT(journal->j_head == journal->j_tail);
2013         J_ASSERT(journal->j_tail_sequence == journal->j_transaction_sequence);
2014         write_unlock(&journal->j_state_lock);
2015 out:
2016         return err;
2017 }
2018
2019 /**
2020  * int jbd2_journal_wipe() - Wipe journal contents
2021  * @journal: Journal to act on.
2022  * @write: flag (see below)
2023  *
2024  * Wipe out all of the contents of a journal, safely.  This will produce
2025  * a warning if the journal contains any valid recovery information.
2026  * Must be called between journal_init_*() and jbd2_journal_load().
2027  *
2028  * If 'write' is non-zero, then we wipe out the journal on disk; otherwise
2029  * we merely suppress recovery.
2030  */
2031
2032 int jbd2_journal_wipe(journal_t *journal, int write)
2033 {
2034         int err = 0;
2035
2036         J_ASSERT (!(journal->j_flags & JBD2_LOADED));
2037
2038         err = load_superblock(journal);
2039         if (err)
2040                 return err;
2041
2042         if (!journal->j_tail)
2043                 goto no_recovery;
2044
2045         printk(KERN_WARNING "JBD2: %s recovery information on journal\n",
2046                 write ? "Clearing" : "Ignoring");
2047
2048         err = jbd2_journal_skip_recovery(journal);
2049         if (write) {
2050                 /* Lock to make assertions happy... */
2051                 mutex_lock(&journal->j_checkpoint_mutex);
2052                 jbd2_mark_journal_empty(journal, WRITE_FUA);
2053                 mutex_unlock(&journal->j_checkpoint_mutex);
2054         }
2055
2056  no_recovery:
2057         return err;
2058 }
2059
2060 /*
2061  * Journal abort has very specific semantics, which we describe
2062  * for journal abort.
2063  *
2064  * Two internal functions, which provide abort to the jbd layer
2065  * itself are here.
2066  */
2067
2068 /*
2069  * Quick version for internal journal use (doesn't lock the journal).
2070  * Aborts hard --- we mark the abort as occurred, but do _nothing_ else,
2071  * and don't attempt to make any other journal updates.
2072  */
2073 void __jbd2_journal_abort_hard(journal_t *journal)
2074 {
2075         transaction_t *transaction;
2076
2077         if (journal->j_flags & JBD2_ABORT)
2078                 return;
2079
2080         printk(KERN_ERR "Aborting journal on device %s.\n",
2081                journal->j_devname);
2082
2083         write_lock(&journal->j_state_lock);
2084         journal->j_flags |= JBD2_ABORT;
2085         transaction = journal->j_running_transaction;
2086         if (transaction)
2087                 __jbd2_log_start_commit(journal, transaction->t_tid);
2088         write_unlock(&journal->j_state_lock);
2089 }
2090
2091 /* Soft abort: record the abort error status in the journal superblock,
2092  * but don't do any other IO. */
2093 static void __journal_abort_soft (journal_t *journal, int errno)
2094 {
2095         if (journal->j_flags & JBD2_ABORT)
2096                 return;
2097
2098         if (!journal->j_errno)
2099                 journal->j_errno = errno;
2100
2101         __jbd2_journal_abort_hard(journal);
2102
2103         jbd2_journal_update_sb_errno(journal);
2104         write_lock(&journal->j_state_lock);
2105         journal->j_flags |= JBD2_REC_ERR;
2106         write_unlock(&journal->j_state_lock);
2107 }
2108
2109 /**
2110  * void jbd2_journal_abort () - Shutdown the journal immediately.
2111  * @journal: the journal to shutdown.
2112  * @errno:   an error number to record in the journal indicating
2113  *           the reason for the shutdown.
2114  *
2115  * Perform a complete, immediate shutdown of the ENTIRE
2116  * journal (not of a single transaction).  This operation cannot be
2117  * undone without closing and reopening the journal.
2118  *
2119  * The jbd2_journal_abort function is intended to support higher level error
2120  * recovery mechanisms such as the ext2/ext3 remount-readonly error
2121  * mode.
2122  *
2123  * Journal abort has very specific semantics.  Any existing dirty,
2124  * unjournaled buffers in the main filesystem will still be written to
2125  * disk by bdflush, but the journaling mechanism will be suspended
2126  * immediately and no further transaction commits will be honoured.
2127  *
2128  * Any dirty, journaled buffers will be written back to disk without
2129  * hitting the journal.  Atomicity cannot be guaranteed on an aborted
2130  * filesystem, but we _do_ attempt to leave as much data as possible
2131  * behind for fsck to use for cleanup.
2132  *
2133  * Any attempt to get a new transaction handle on a journal which is in
2134  * ABORT state will just result in an -EROFS error return.  A
2135  * jbd2_journal_stop on an existing handle will return -EIO if we have
2136  * entered abort state during the update.
2137  *
2138  * Recursive transactions are not disturbed by journal abort until the
2139  * final jbd2_journal_stop, which will receive the -EIO error.
2140  *
2141  * Finally, the jbd2_journal_abort call allows the caller to supply an errno
2142  * which will be recorded (if possible) in the journal superblock.  This
2143  * allows a client to record failure conditions in the middle of a
2144  * transaction without having to complete the transaction to record the
2145  * failure to disk.  ext3_error, for example, now uses this
2146  * functionality.
2147  *
2148  */
2149
2150 void jbd2_journal_abort(journal_t *journal, int errno)
2151 {
2152         __journal_abort_soft(journal, errno);
2153 }
2154
2155 /**
2156  * int jbd2_journal_errno () - returns the journal's error state.
2157  * @journal: journal to examine.
2158  *
2159  * This is the errno number set with jbd2_journal_abort(), the last
2160  * time the journal was mounted - if the journal was stopped
2161  * without calling abort this will be 0.
2162  *
2163  * If the journal has been aborted on this mount time -EROFS will
2164  * be returned.
2165  */
2166 int jbd2_journal_errno(journal_t *journal)
2167 {
2168         int err;
2169
2170         read_lock(&journal->j_state_lock);
2171         if (journal->j_flags & JBD2_ABORT)
2172                 err = -EROFS;
2173         else
2174                 err = journal->j_errno;
2175         read_unlock(&journal->j_state_lock);
2176         return err;
2177 }
2178
2179 /**
2180  * int jbd2_journal_clear_err () - clears the journal's error state
2181  * @journal: journal to act on.
2182  *
2183  * An error must be cleared or acked to take a FS out of readonly
2184  * mode.
2185  */
2186 int jbd2_journal_clear_err(journal_t *journal)
2187 {
2188         int err = 0;
2189
2190         write_lock(&journal->j_state_lock);
2191         if (journal->j_flags & JBD2_ABORT)
2192                 err = -EROFS;
2193         else
2194                 journal->j_errno = 0;
2195         write_unlock(&journal->j_state_lock);
2196         return err;
2197 }
2198
2199 /**
2200  * void jbd2_journal_ack_err() - Ack journal err.
2201  * @journal: journal to act on.
2202  *
2203  * An error must be cleared or acked to take a FS out of readonly
2204  * mode.
2205  */
2206 void jbd2_journal_ack_err(journal_t *journal)
2207 {
2208         write_lock(&journal->j_state_lock);
2209         if (journal->j_errno)
2210                 journal->j_flags |= JBD2_ACK_ERR;
2211         write_unlock(&journal->j_state_lock);
2212 }
2213
2214 int jbd2_journal_blocks_per_page(struct inode *inode)
2215 {
2216         return 1 << (PAGE_SHIFT - inode->i_sb->s_blocksize_bits);
2217 }
2218
2219 /*
2220  * helper functions to deal with 32 or 64bit block numbers.
2221  */
2222 size_t journal_tag_bytes(journal_t *journal)
2223 {
2224         size_t sz;
2225
2226         if (jbd2_has_feature_csum3(journal))
2227                 return sizeof(journal_block_tag3_t);
2228
2229         sz = sizeof(journal_block_tag_t);
2230
2231         if (jbd2_has_feature_csum2(journal))
2232                 sz += sizeof(__u16);
2233
2234         if (jbd2_has_feature_64bit(journal))
2235                 return sz;
2236         else
2237                 return sz - sizeof(__u32);
2238 }
2239
2240 /*
2241  * JBD memory management
2242  *
2243  * These functions are used to allocate block-sized chunks of memory
2244  * used for making copies of buffer_head data.  Very often it will be
2245  * page-sized chunks of data, but sometimes it will be in
2246  * sub-page-size chunks.  (For example, 16k pages on Power systems
2247  * with a 4k block file system.)  For blocks smaller than a page, we
2248  * use a SLAB allocator.  There are slab caches for each block size,
2249  * which are allocated at mount time, if necessary, and we only free
2250  * (all of) the slab caches when/if the jbd2 module is unloaded.  For
2251  * this reason we don't need to a mutex to protect access to
2252  * jbd2_slab[] allocating or releasing memory; only in
2253  * jbd2_journal_create_slab().
2254  */
2255 #define JBD2_MAX_SLABS 8
2256 static struct kmem_cache *jbd2_slab[JBD2_MAX_SLABS];
2257
2258 static const char *jbd2_slab_names[JBD2_MAX_SLABS] = {
2259         "jbd2_1k", "jbd2_2k", "jbd2_4k", "jbd2_8k",
2260         "jbd2_16k", "jbd2_32k", "jbd2_64k", "jbd2_128k"
2261 };
2262
2263
2264 static void jbd2_journal_destroy_slabs(void)
2265 {
2266         int i;
2267
2268         for (i = 0; i < JBD2_MAX_SLABS; i++) {
2269                 if (jbd2_slab[i])
2270                         kmem_cache_destroy(jbd2_slab[i]);
2271                 jbd2_slab[i] = NULL;
2272         }
2273 }
2274
2275 static int jbd2_journal_create_slab(size_t size)
2276 {
2277         static DEFINE_MUTEX(jbd2_slab_create_mutex);
2278         int i = order_base_2(size) - 10;
2279         size_t slab_size;
2280
2281         if (size == PAGE_SIZE)
2282                 return 0;
2283
2284         if (i >= JBD2_MAX_SLABS)
2285                 return -EINVAL;
2286
2287         if (unlikely(i < 0))
2288                 i = 0;
2289         mutex_lock(&jbd2_slab_create_mutex);
2290         if (jbd2_slab[i]) {
2291                 mutex_unlock(&jbd2_slab_create_mutex);
2292                 return 0;       /* Already created */
2293         }
2294
2295         slab_size = 1 << (i+10);
2296         jbd2_slab[i] = kmem_cache_create(jbd2_slab_names[i], slab_size,
2297                                          slab_size, 0, NULL);
2298         mutex_unlock(&jbd2_slab_create_mutex);
2299         if (!jbd2_slab[i]) {
2300                 printk(KERN_EMERG "JBD2: no memory for jbd2_slab cache\n");
2301                 return -ENOMEM;
2302         }
2303         return 0;
2304 }
2305
2306 static struct kmem_cache *get_slab(size_t size)
2307 {
2308         int i = order_base_2(size) - 10;
2309
2310         BUG_ON(i >= JBD2_MAX_SLABS);
2311         if (unlikely(i < 0))
2312                 i = 0;
2313         BUG_ON(jbd2_slab[i] == NULL);
2314         return jbd2_slab[i];
2315 }
2316
2317 void *jbd2_alloc(size_t size, gfp_t flags)
2318 {
2319         void *ptr;
2320
2321         BUG_ON(size & (size-1)); /* Must be a power of 2 */
2322
2323         if (size < PAGE_SIZE)
2324                 ptr = kmem_cache_alloc(get_slab(size), flags);
2325         else
2326                 ptr = (void *)__get_free_pages(flags, get_order(size));
2327
2328         /* Check alignment; SLUB has gotten this wrong in the past,
2329          * and this can lead to user data corruption! */
2330         BUG_ON(((unsigned long) ptr) & (size-1));
2331
2332         return ptr;
2333 }
2334
2335 void jbd2_free(void *ptr, size_t size)
2336 {
2337         if (size < PAGE_SIZE)
2338                 kmem_cache_free(get_slab(size), ptr);
2339         else
2340                 free_pages((unsigned long)ptr, get_order(size));
2341 };
2342
2343 /*
2344  * Journal_head storage management
2345  */
2346 static struct kmem_cache *jbd2_journal_head_cache;
2347 #ifdef CONFIG_JBD2_DEBUG
2348 static atomic_t nr_journal_heads = ATOMIC_INIT(0);
2349 #endif
2350
2351 static int jbd2_journal_init_journal_head_cache(void)
2352 {
2353         int retval;
2354
2355         J_ASSERT(jbd2_journal_head_cache == NULL);
2356         jbd2_journal_head_cache = kmem_cache_create("jbd2_journal_head",
2357                                 sizeof(struct journal_head),
2358                                 0,              /* offset */
2359                                 SLAB_TEMPORARY | SLAB_DESTROY_BY_RCU,
2360                                 NULL);          /* ctor */
2361         retval = 0;
2362         if (!jbd2_journal_head_cache) {
2363                 retval = -ENOMEM;
2364                 printk(KERN_EMERG "JBD2: no memory for journal_head cache\n");
2365         }
2366         return retval;
2367 }
2368
2369 static void jbd2_journal_destroy_journal_head_cache(void)
2370 {
2371         if (jbd2_journal_head_cache) {
2372                 kmem_cache_destroy(jbd2_journal_head_cache);
2373                 jbd2_journal_head_cache = NULL;
2374         }
2375 }
2376
2377 /*
2378  * journal_head splicing and dicing
2379  */
2380 static struct journal_head *journal_alloc_journal_head(void)
2381 {
2382         struct journal_head *ret;
2383
2384 #ifdef CONFIG_JBD2_DEBUG
2385         atomic_inc(&nr_journal_heads);
2386 #endif
2387         ret = kmem_cache_zalloc(jbd2_journal_head_cache, GFP_NOFS);
2388         if (!ret) {
2389                 jbd_debug(1, "out of memory for journal_head\n");
2390                 pr_notice_ratelimited("ENOMEM in %s, retrying.\n", __func__);
2391                 ret = kmem_cache_zalloc(jbd2_journal_head_cache,
2392                                 GFP_NOFS | __GFP_NOFAIL);
2393         }
2394         return ret;
2395 }
2396
2397 static void journal_free_journal_head(struct journal_head *jh)
2398 {
2399 #ifdef CONFIG_JBD2_DEBUG
2400         atomic_dec(&nr_journal_heads);
2401         memset(jh, JBD2_POISON_FREE, sizeof(*jh));
2402 #endif
2403         kmem_cache_free(jbd2_journal_head_cache, jh);
2404 }
2405
2406 /*
2407  * A journal_head is attached to a buffer_head whenever JBD has an
2408  * interest in the buffer.
2409  *
2410  * Whenever a buffer has an attached journal_head, its ->b_state:BH_JBD bit
2411  * is set.  This bit is tested in core kernel code where we need to take
2412  * JBD-specific actions.  Testing the zeroness of ->b_private is not reliable
2413  * there.
2414  *
2415  * When a buffer has its BH_JBD bit set, its ->b_count is elevated by one.
2416  *
2417  * When a buffer has its BH_JBD bit set it is immune from being released by
2418  * core kernel code, mainly via ->b_count.
2419  *
2420  * A journal_head is detached from its buffer_head when the journal_head's
2421  * b_jcount reaches zero. Running transaction (b_transaction) and checkpoint
2422  * transaction (b_cp_transaction) hold their references to b_jcount.
2423  *
2424  * Various places in the kernel want to attach a journal_head to a buffer_head
2425  * _before_ attaching the journal_head to a transaction.  To protect the
2426  * journal_head in this situation, jbd2_journal_add_journal_head elevates the
2427  * journal_head's b_jcount refcount by one.  The caller must call
2428  * jbd2_journal_put_journal_head() to undo this.
2429  *
2430  * So the typical usage would be:
2431  *
2432  *      (Attach a journal_head if needed.  Increments b_jcount)
2433  *      struct journal_head *jh = jbd2_journal_add_journal_head(bh);
2434  *      ...
2435  *      (Get another reference for transaction)
2436  *      jbd2_journal_grab_journal_head(bh);
2437  *      jh->b_transaction = xxx;
2438  *      (Put original reference)
2439  *      jbd2_journal_put_journal_head(jh);
2440  */
2441
2442 /*
2443  * Give a buffer_head a journal_head.
2444  *
2445  * May sleep.
2446  */
2447 struct journal_head *jbd2_journal_add_journal_head(struct buffer_head *bh)
2448 {
2449         struct journal_head *jh;
2450         struct journal_head *new_jh = NULL;
2451
2452 repeat:
2453         if (!buffer_jbd(bh))
2454                 new_jh = journal_alloc_journal_head();
2455
2456         jbd_lock_bh_journal_head(bh);
2457         if (buffer_jbd(bh)) {
2458                 jh = bh2jh(bh);
2459         } else {
2460                 J_ASSERT_BH(bh,
2461                         (atomic_read(&bh->b_count) > 0) ||
2462                         (bh->b_page && bh->b_page->mapping));
2463
2464                 if (!new_jh) {
2465                         jbd_unlock_bh_journal_head(bh);
2466                         goto repeat;
2467                 }
2468
2469                 jh = new_jh;
2470                 new_jh = NULL;          /* We consumed it */
2471                 set_buffer_jbd(bh);
2472                 bh->b_private = jh;
2473                 jh->b_bh = bh;
2474                 get_bh(bh);
2475                 BUFFER_TRACE(bh, "added journal_head");
2476         }
2477         jh->b_jcount++;
2478         jbd_unlock_bh_journal_head(bh);
2479         if (new_jh)
2480                 journal_free_journal_head(new_jh);
2481         return bh->b_private;
2482 }
2483
2484 /*
2485  * Grab a ref against this buffer_head's journal_head.  If it ended up not
2486  * having a journal_head, return NULL
2487  */
2488 struct journal_head *jbd2_journal_grab_journal_head(struct buffer_head *bh)
2489 {
2490         struct journal_head *jh = NULL;
2491
2492         jbd_lock_bh_journal_head(bh);
2493         if (buffer_jbd(bh)) {
2494                 jh = bh2jh(bh);
2495                 jh->b_jcount++;
2496         }
2497         jbd_unlock_bh_journal_head(bh);
2498         return jh;
2499 }
2500
2501 static void __journal_remove_journal_head(struct buffer_head *bh)
2502 {
2503         struct journal_head *jh = bh2jh(bh);
2504
2505         J_ASSERT_JH(jh, jh->b_jcount >= 0);
2506         J_ASSERT_JH(jh, jh->b_transaction == NULL);
2507         J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
2508         J_ASSERT_JH(jh, jh->b_cp_transaction == NULL);
2509         J_ASSERT_JH(jh, jh->b_jlist == BJ_None);
2510         J_ASSERT_BH(bh, buffer_jbd(bh));
2511         J_ASSERT_BH(bh, jh2bh(jh) == bh);
2512         BUFFER_TRACE(bh, "remove journal_head");
2513         if (jh->b_frozen_data) {
2514                 printk(KERN_WARNING "%s: freeing b_frozen_data\n", __func__);
2515                 jbd2_free(jh->b_frozen_data, bh->b_size);
2516         }
2517         if (jh->b_committed_data) {
2518                 printk(KERN_WARNING "%s: freeing b_committed_data\n", __func__);
2519                 jbd2_free(jh->b_committed_data, bh->b_size);
2520         }
2521         bh->b_private = NULL;
2522         jh->b_bh = NULL;        /* debug, really */
2523         clear_buffer_jbd(bh);
2524         journal_free_journal_head(jh);
2525 }
2526
2527 /*
2528  * Drop a reference on the passed journal_head.  If it fell to zero then
2529  * release the journal_head from the buffer_head.
2530  */
2531 void jbd2_journal_put_journal_head(struct journal_head *jh)
2532 {
2533         struct buffer_head *bh = jh2bh(jh);
2534
2535         jbd_lock_bh_journal_head(bh);
2536         J_ASSERT_JH(jh, jh->b_jcount > 0);
2537         --jh->b_jcount;
2538         if (!jh->b_jcount) {
2539                 __journal_remove_journal_head(bh);
2540                 jbd_unlock_bh_journal_head(bh);
2541                 __brelse(bh);
2542         } else
2543                 jbd_unlock_bh_journal_head(bh);
2544 }
2545
2546 /*
2547  * Initialize jbd inode head
2548  */
2549 void jbd2_journal_init_jbd_inode(struct jbd2_inode *jinode, struct inode *inode)
2550 {
2551         jinode->i_transaction = NULL;
2552         jinode->i_next_transaction = NULL;
2553         jinode->i_vfs_inode = inode;
2554         jinode->i_flags = 0;
2555         INIT_LIST_HEAD(&jinode->i_list);
2556 }
2557
2558 /*
2559  * Function to be called before we start removing inode from memory (i.e.,
2560  * clear_inode() is a fine place to be called from). It removes inode from
2561  * transaction's lists.
2562  */
2563 void jbd2_journal_release_jbd_inode(journal_t *journal,
2564                                     struct jbd2_inode *jinode)
2565 {
2566         if (!journal)
2567                 return;
2568 restart:
2569         spin_lock(&journal->j_list_lock);
2570         /* Is commit writing out inode - we have to wait */
2571         if (jinode->i_flags & JI_COMMIT_RUNNING) {
2572                 wait_queue_head_t *wq;
2573                 DEFINE_WAIT_BIT(wait, &jinode->i_flags, __JI_COMMIT_RUNNING);
2574                 wq = bit_waitqueue(&jinode->i_flags, __JI_COMMIT_RUNNING);
2575                 prepare_to_wait(wq, &wait.wait, TASK_UNINTERRUPTIBLE);
2576                 spin_unlock(&journal->j_list_lock);
2577                 schedule();
2578                 finish_wait(wq, &wait.wait);
2579                 goto restart;
2580         }
2581
2582         if (jinode->i_transaction) {
2583                 list_del(&jinode->i_list);
2584                 jinode->i_transaction = NULL;
2585         }
2586         spin_unlock(&journal->j_list_lock);
2587 }
2588
2589
2590 #ifdef CONFIG_PROC_FS
2591
2592 #define JBD2_STATS_PROC_NAME "fs/jbd2"
2593
2594 static void __init jbd2_create_jbd_stats_proc_entry(void)
2595 {
2596         proc_jbd2_stats = proc_mkdir(JBD2_STATS_PROC_NAME, NULL);
2597 }
2598
2599 static void __exit jbd2_remove_jbd_stats_proc_entry(void)
2600 {
2601         if (proc_jbd2_stats)
2602                 remove_proc_entry(JBD2_STATS_PROC_NAME, NULL);
2603 }
2604
2605 #else
2606
2607 #define jbd2_create_jbd_stats_proc_entry() do {} while (0)
2608 #define jbd2_remove_jbd_stats_proc_entry() do {} while (0)
2609
2610 #endif
2611
2612 struct kmem_cache *jbd2_handle_cache, *jbd2_inode_cache;
2613
2614 static int __init jbd2_journal_init_handle_cache(void)
2615 {
2616         jbd2_handle_cache = KMEM_CACHE(jbd2_journal_handle, SLAB_TEMPORARY);
2617         if (jbd2_handle_cache == NULL) {
2618                 printk(KERN_EMERG "JBD2: failed to create handle cache\n");
2619                 return -ENOMEM;
2620         }
2621         jbd2_inode_cache = KMEM_CACHE(jbd2_inode, 0);
2622         if (jbd2_inode_cache == NULL) {
2623                 printk(KERN_EMERG "JBD2: failed to create inode cache\n");
2624                 kmem_cache_destroy(jbd2_handle_cache);
2625                 return -ENOMEM;
2626         }
2627         return 0;
2628 }
2629
2630 static void jbd2_journal_destroy_handle_cache(void)
2631 {
2632         if (jbd2_handle_cache)
2633                 kmem_cache_destroy(jbd2_handle_cache);
2634         if (jbd2_inode_cache)
2635                 kmem_cache_destroy(jbd2_inode_cache);
2636
2637 }
2638
2639 /*
2640  * Module startup and shutdown
2641  */
2642
2643 static int __init journal_init_caches(void)
2644 {
2645         int ret;
2646
2647         ret = jbd2_journal_init_revoke_caches();
2648         if (ret == 0)
2649                 ret = jbd2_journal_init_journal_head_cache();
2650         if (ret == 0)
2651                 ret = jbd2_journal_init_handle_cache();
2652         if (ret == 0)
2653                 ret = jbd2_journal_init_transaction_cache();
2654         return ret;
2655 }
2656
2657 static void jbd2_journal_destroy_caches(void)
2658 {
2659         jbd2_journal_destroy_revoke_caches();
2660         jbd2_journal_destroy_journal_head_cache();
2661         jbd2_journal_destroy_handle_cache();
2662         jbd2_journal_destroy_transaction_cache();
2663         jbd2_journal_destroy_slabs();
2664 }
2665
2666 static int __init journal_init(void)
2667 {
2668         int ret;
2669
2670         BUILD_BUG_ON(sizeof(struct journal_superblock_s) != 1024);
2671
2672         ret = journal_init_caches();
2673         if (ret == 0) {
2674                 jbd2_create_jbd_stats_proc_entry();
2675         } else {
2676                 jbd2_journal_destroy_caches();
2677         }
2678         return ret;
2679 }
2680
2681 static void __exit journal_exit(void)
2682 {
2683 #ifdef CONFIG_JBD2_DEBUG
2684         int n = atomic_read(&nr_journal_heads);
2685         if (n)
2686                 printk(KERN_ERR "JBD2: leaked %d journal_heads!\n", n);
2687 #endif
2688         jbd2_remove_jbd_stats_proc_entry();
2689         jbd2_journal_destroy_caches();
2690 }
2691
2692 MODULE_LICENSE("GPL");
2693 module_init(journal_init);
2694 module_exit(journal_exit);
2695