jbd2: cleanup write flags handling from jbd2_write_superblock()
[sfrench/cifs-2.6.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 <linux/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,
930                                               REQ_SYNC | REQ_FUA);
931         if (ret)
932                 goto out;
933
934         write_lock(&journal->j_state_lock);
935         freed = block - journal->j_tail;
936         if (block < journal->j_tail)
937                 freed += journal->j_last - journal->j_first;
938
939         trace_jbd2_update_log_tail(journal, tid, block, freed);
940         jbd_debug(1,
941                   "Cleaning journal tail from %d to %d (offset %lu), "
942                   "freeing %lu\n",
943                   journal->j_tail_sequence, tid, block, freed);
944
945         journal->j_free += freed;
946         journal->j_tail_sequence = tid;
947         journal->j_tail = block;
948         write_unlock(&journal->j_state_lock);
949
950 out:
951         return ret;
952 }
953
954 /*
955  * This is a variaon of __jbd2_update_log_tail which checks for validity of
956  * provided log tail and locks j_checkpoint_mutex. So it is safe against races
957  * with other threads updating log tail.
958  */
959 void jbd2_update_log_tail(journal_t *journal, tid_t tid, unsigned long block)
960 {
961         mutex_lock_io(&journal->j_checkpoint_mutex);
962         if (tid_gt(tid, journal->j_tail_sequence))
963                 __jbd2_update_log_tail(journal, tid, block);
964         mutex_unlock(&journal->j_checkpoint_mutex);
965 }
966
967 struct jbd2_stats_proc_session {
968         journal_t *journal;
969         struct transaction_stats_s *stats;
970         int start;
971         int max;
972 };
973
974 static void *jbd2_seq_info_start(struct seq_file *seq, loff_t *pos)
975 {
976         return *pos ? NULL : SEQ_START_TOKEN;
977 }
978
979 static void *jbd2_seq_info_next(struct seq_file *seq, void *v, loff_t *pos)
980 {
981         return NULL;
982 }
983
984 static int jbd2_seq_info_show(struct seq_file *seq, void *v)
985 {
986         struct jbd2_stats_proc_session *s = seq->private;
987
988         if (v != SEQ_START_TOKEN)
989                 return 0;
990         seq_printf(seq, "%lu transactions (%lu requested), "
991                    "each up to %u blocks\n",
992                    s->stats->ts_tid, s->stats->ts_requested,
993                    s->journal->j_max_transaction_buffers);
994         if (s->stats->ts_tid == 0)
995                 return 0;
996         seq_printf(seq, "average: \n  %ums waiting for transaction\n",
997             jiffies_to_msecs(s->stats->run.rs_wait / s->stats->ts_tid));
998         seq_printf(seq, "  %ums request delay\n",
999             (s->stats->ts_requested == 0) ? 0 :
1000             jiffies_to_msecs(s->stats->run.rs_request_delay /
1001                              s->stats->ts_requested));
1002         seq_printf(seq, "  %ums running transaction\n",
1003             jiffies_to_msecs(s->stats->run.rs_running / s->stats->ts_tid));
1004         seq_printf(seq, "  %ums transaction was being locked\n",
1005             jiffies_to_msecs(s->stats->run.rs_locked / s->stats->ts_tid));
1006         seq_printf(seq, "  %ums flushing data (in ordered mode)\n",
1007             jiffies_to_msecs(s->stats->run.rs_flushing / s->stats->ts_tid));
1008         seq_printf(seq, "  %ums logging transaction\n",
1009             jiffies_to_msecs(s->stats->run.rs_logging / s->stats->ts_tid));
1010         seq_printf(seq, "  %lluus average transaction commit time\n",
1011                    div_u64(s->journal->j_average_commit_time, 1000));
1012         seq_printf(seq, "  %lu handles per transaction\n",
1013             s->stats->run.rs_handle_count / s->stats->ts_tid);
1014         seq_printf(seq, "  %lu blocks per transaction\n",
1015             s->stats->run.rs_blocks / s->stats->ts_tid);
1016         seq_printf(seq, "  %lu logged blocks per transaction\n",
1017             s->stats->run.rs_blocks_logged / s->stats->ts_tid);
1018         return 0;
1019 }
1020
1021 static void jbd2_seq_info_stop(struct seq_file *seq, void *v)
1022 {
1023 }
1024
1025 static const struct seq_operations jbd2_seq_info_ops = {
1026         .start  = jbd2_seq_info_start,
1027         .next   = jbd2_seq_info_next,
1028         .stop   = jbd2_seq_info_stop,
1029         .show   = jbd2_seq_info_show,
1030 };
1031
1032 static int jbd2_seq_info_open(struct inode *inode, struct file *file)
1033 {
1034         journal_t *journal = PDE_DATA(inode);
1035         struct jbd2_stats_proc_session *s;
1036         int rc, size;
1037
1038         s = kmalloc(sizeof(*s), GFP_KERNEL);
1039         if (s == NULL)
1040                 return -ENOMEM;
1041         size = sizeof(struct transaction_stats_s);
1042         s->stats = kmalloc(size, GFP_KERNEL);
1043         if (s->stats == NULL) {
1044                 kfree(s);
1045                 return -ENOMEM;
1046         }
1047         spin_lock(&journal->j_history_lock);
1048         memcpy(s->stats, &journal->j_stats, size);
1049         s->journal = journal;
1050         spin_unlock(&journal->j_history_lock);
1051
1052         rc = seq_open(file, &jbd2_seq_info_ops);
1053         if (rc == 0) {
1054                 struct seq_file *m = file->private_data;
1055                 m->private = s;
1056         } else {
1057                 kfree(s->stats);
1058                 kfree(s);
1059         }
1060         return rc;
1061
1062 }
1063
1064 static int jbd2_seq_info_release(struct inode *inode, struct file *file)
1065 {
1066         struct seq_file *seq = file->private_data;
1067         struct jbd2_stats_proc_session *s = seq->private;
1068         kfree(s->stats);
1069         kfree(s);
1070         return seq_release(inode, file);
1071 }
1072
1073 static const struct file_operations jbd2_seq_info_fops = {
1074         .owner          = THIS_MODULE,
1075         .open           = jbd2_seq_info_open,
1076         .read           = seq_read,
1077         .llseek         = seq_lseek,
1078         .release        = jbd2_seq_info_release,
1079 };
1080
1081 static struct proc_dir_entry *proc_jbd2_stats;
1082
1083 static void jbd2_stats_proc_init(journal_t *journal)
1084 {
1085         journal->j_proc_entry = proc_mkdir(journal->j_devname, proc_jbd2_stats);
1086         if (journal->j_proc_entry) {
1087                 proc_create_data("info", S_IRUGO, journal->j_proc_entry,
1088                                  &jbd2_seq_info_fops, journal);
1089         }
1090 }
1091
1092 static void jbd2_stats_proc_exit(journal_t *journal)
1093 {
1094         remove_proc_entry("info", journal->j_proc_entry);
1095         remove_proc_entry(journal->j_devname, proc_jbd2_stats);
1096 }
1097
1098 /*
1099  * Management for journal control blocks: functions to create and
1100  * destroy journal_t structures, and to initialise and read existing
1101  * journal blocks from disk.  */
1102
1103 /* First: create and setup a journal_t object in memory.  We initialise
1104  * very few fields yet: that has to wait until we have created the
1105  * journal structures from from scratch, or loaded them from disk. */
1106
1107 static journal_t *journal_init_common(struct block_device *bdev,
1108                         struct block_device *fs_dev,
1109                         unsigned long long start, int len, int blocksize)
1110 {
1111         static struct lock_class_key jbd2_trans_commit_key;
1112         journal_t *journal;
1113         int err;
1114         struct buffer_head *bh;
1115         int n;
1116
1117         journal = kzalloc(sizeof(*journal), GFP_KERNEL);
1118         if (!journal)
1119                 return NULL;
1120
1121         init_waitqueue_head(&journal->j_wait_transaction_locked);
1122         init_waitqueue_head(&journal->j_wait_done_commit);
1123         init_waitqueue_head(&journal->j_wait_commit);
1124         init_waitqueue_head(&journal->j_wait_updates);
1125         init_waitqueue_head(&journal->j_wait_reserved);
1126         mutex_init(&journal->j_barrier);
1127         mutex_init(&journal->j_checkpoint_mutex);
1128         spin_lock_init(&journal->j_revoke_lock);
1129         spin_lock_init(&journal->j_list_lock);
1130         rwlock_init(&journal->j_state_lock);
1131
1132         journal->j_commit_interval = (HZ * JBD2_DEFAULT_MAX_COMMIT_AGE);
1133         journal->j_min_batch_time = 0;
1134         journal->j_max_batch_time = 15000; /* 15ms */
1135         atomic_set(&journal->j_reserved_credits, 0);
1136
1137         /* The journal is marked for error until we succeed with recovery! */
1138         journal->j_flags = JBD2_ABORT;
1139
1140         /* Set up a default-sized revoke table for the new mount. */
1141         err = jbd2_journal_init_revoke(journal, JOURNAL_REVOKE_DEFAULT_HASH);
1142         if (err)
1143                 goto err_cleanup;
1144
1145         spin_lock_init(&journal->j_history_lock);
1146
1147         lockdep_init_map(&journal->j_trans_commit_map, "jbd2_handle",
1148                          &jbd2_trans_commit_key, 0);
1149
1150         /* journal descriptor can store up to n blocks -bzzz */
1151         journal->j_blocksize = blocksize;
1152         journal->j_dev = bdev;
1153         journal->j_fs_dev = fs_dev;
1154         journal->j_blk_offset = start;
1155         journal->j_maxlen = len;
1156         n = journal->j_blocksize / sizeof(journal_block_tag_t);
1157         journal->j_wbufsize = n;
1158         journal->j_wbuf = kmalloc_array(n, sizeof(struct buffer_head *),
1159                                         GFP_KERNEL);
1160         if (!journal->j_wbuf)
1161                 goto err_cleanup;
1162
1163         bh = getblk_unmovable(journal->j_dev, start, journal->j_blocksize);
1164         if (!bh) {
1165                 pr_err("%s: Cannot get buffer for journal superblock\n",
1166                         __func__);
1167                 goto err_cleanup;
1168         }
1169         journal->j_sb_buffer = bh;
1170         journal->j_superblock = (journal_superblock_t *)bh->b_data;
1171
1172         return journal;
1173
1174 err_cleanup:
1175         kfree(journal->j_wbuf);
1176         jbd2_journal_destroy_revoke(journal);
1177         kfree(journal);
1178         return NULL;
1179 }
1180
1181 /* jbd2_journal_init_dev and jbd2_journal_init_inode:
1182  *
1183  * Create a journal structure assigned some fixed set of disk blocks to
1184  * the journal.  We don't actually touch those disk blocks yet, but we
1185  * need to set up all of the mapping information to tell the journaling
1186  * system where the journal blocks are.
1187  *
1188  */
1189
1190 /**
1191  *  journal_t * jbd2_journal_init_dev() - creates and initialises a journal structure
1192  *  @bdev: Block device on which to create the journal
1193  *  @fs_dev: Device which hold journalled filesystem for this journal.
1194  *  @start: Block nr Start of journal.
1195  *  @len:  Length of the journal in blocks.
1196  *  @blocksize: blocksize of journalling device
1197  *
1198  *  Returns: a newly created journal_t *
1199  *
1200  *  jbd2_journal_init_dev creates a journal which maps a fixed contiguous
1201  *  range of blocks on an arbitrary block device.
1202  *
1203  */
1204 journal_t *jbd2_journal_init_dev(struct block_device *bdev,
1205                         struct block_device *fs_dev,
1206                         unsigned long long start, int len, int blocksize)
1207 {
1208         journal_t *journal;
1209
1210         journal = journal_init_common(bdev, fs_dev, start, len, blocksize);
1211         if (!journal)
1212                 return NULL;
1213
1214         bdevname(journal->j_dev, journal->j_devname);
1215         strreplace(journal->j_devname, '/', '!');
1216         jbd2_stats_proc_init(journal);
1217
1218         return journal;
1219 }
1220
1221 /**
1222  *  journal_t * jbd2_journal_init_inode () - creates a journal which maps to a inode.
1223  *  @inode: An inode to create the journal in
1224  *
1225  * jbd2_journal_init_inode creates a journal which maps an on-disk inode as
1226  * the journal.  The inode must exist already, must support bmap() and
1227  * must have all data blocks preallocated.
1228  */
1229 journal_t *jbd2_journal_init_inode(struct inode *inode)
1230 {
1231         journal_t *journal;
1232         char *p;
1233         unsigned long long blocknr;
1234
1235         blocknr = bmap(inode, 0);
1236         if (!blocknr) {
1237                 pr_err("%s: Cannot locate journal superblock\n",
1238                         __func__);
1239                 return NULL;
1240         }
1241
1242         jbd_debug(1, "JBD2: inode %s/%ld, size %lld, bits %d, blksize %ld\n",
1243                   inode->i_sb->s_id, inode->i_ino, (long long) inode->i_size,
1244                   inode->i_sb->s_blocksize_bits, inode->i_sb->s_blocksize);
1245
1246         journal = journal_init_common(inode->i_sb->s_bdev, inode->i_sb->s_bdev,
1247                         blocknr, inode->i_size >> inode->i_sb->s_blocksize_bits,
1248                         inode->i_sb->s_blocksize);
1249         if (!journal)
1250                 return NULL;
1251
1252         journal->j_inode = inode;
1253         bdevname(journal->j_dev, journal->j_devname);
1254         p = strreplace(journal->j_devname, '/', '!');
1255         sprintf(p, "-%lu", journal->j_inode->i_ino);
1256         jbd2_stats_proc_init(journal);
1257
1258         return journal;
1259 }
1260
1261 /*
1262  * If the journal init or create aborts, we need to mark the journal
1263  * superblock as being NULL to prevent the journal destroy from writing
1264  * back a bogus superblock.
1265  */
1266 static void journal_fail_superblock (journal_t *journal)
1267 {
1268         struct buffer_head *bh = journal->j_sb_buffer;
1269         brelse(bh);
1270         journal->j_sb_buffer = NULL;
1271 }
1272
1273 /*
1274  * Given a journal_t structure, initialise the various fields for
1275  * startup of a new journaling session.  We use this both when creating
1276  * a journal, and after recovering an old journal to reset it for
1277  * subsequent use.
1278  */
1279
1280 static int journal_reset(journal_t *journal)
1281 {
1282         journal_superblock_t *sb = journal->j_superblock;
1283         unsigned long long first, last;
1284
1285         first = be32_to_cpu(sb->s_first);
1286         last = be32_to_cpu(sb->s_maxlen);
1287         if (first + JBD2_MIN_JOURNAL_BLOCKS > last + 1) {
1288                 printk(KERN_ERR "JBD2: Journal too short (blocks %llu-%llu).\n",
1289                        first, last);
1290                 journal_fail_superblock(journal);
1291                 return -EINVAL;
1292         }
1293
1294         journal->j_first = first;
1295         journal->j_last = last;
1296
1297         journal->j_head = first;
1298         journal->j_tail = first;
1299         journal->j_free = last - first;
1300
1301         journal->j_tail_sequence = journal->j_transaction_sequence;
1302         journal->j_commit_sequence = journal->j_transaction_sequence - 1;
1303         journal->j_commit_request = journal->j_commit_sequence;
1304
1305         journal->j_max_transaction_buffers = journal->j_maxlen / 4;
1306
1307         /*
1308          * As a special case, if the on-disk copy is already marked as needing
1309          * no recovery (s_start == 0), then we can safely defer the superblock
1310          * update until the next commit by setting JBD2_FLUSHED.  This avoids
1311          * attempting a write to a potential-readonly device.
1312          */
1313         if (sb->s_start == 0) {
1314                 jbd_debug(1, "JBD2: Skipping superblock update on recovered sb "
1315                         "(start %ld, seq %d, errno %d)\n",
1316                         journal->j_tail, journal->j_tail_sequence,
1317                         journal->j_errno);
1318                 journal->j_flags |= JBD2_FLUSHED;
1319         } else {
1320                 /* Lock here to make assertions happy... */
1321                 mutex_lock_io(&journal->j_checkpoint_mutex);
1322                 /*
1323                  * Update log tail information. We use REQ_FUA since new
1324                  * transaction will start reusing journal space and so we
1325                  * must make sure information about current log tail is on
1326                  * disk before that.
1327                  */
1328                 jbd2_journal_update_sb_log_tail(journal,
1329                                                 journal->j_tail_sequence,
1330                                                 journal->j_tail,
1331                                                 REQ_SYNC | REQ_FUA);
1332                 mutex_unlock(&journal->j_checkpoint_mutex);
1333         }
1334         return jbd2_journal_start_thread(journal);
1335 }
1336
1337 static int jbd2_write_superblock(journal_t *journal, int write_flags)
1338 {
1339         struct buffer_head *bh = journal->j_sb_buffer;
1340         journal_superblock_t *sb = journal->j_superblock;
1341         int ret;
1342
1343         trace_jbd2_write_superblock(journal, write_flags);
1344         if (!(journal->j_flags & JBD2_BARRIER))
1345                 write_flags &= ~(REQ_FUA | REQ_PREFLUSH);
1346         lock_buffer(bh);
1347         if (buffer_write_io_error(bh)) {
1348                 /*
1349                  * Oh, dear.  A previous attempt to write the journal
1350                  * superblock failed.  This could happen because the
1351                  * USB device was yanked out.  Or it could happen to
1352                  * be a transient write error and maybe the block will
1353                  * be remapped.  Nothing we can do but to retry the
1354                  * write and hope for the best.
1355                  */
1356                 printk(KERN_ERR "JBD2: previous I/O error detected "
1357                        "for journal superblock update for %s.\n",
1358                        journal->j_devname);
1359                 clear_buffer_write_io_error(bh);
1360                 set_buffer_uptodate(bh);
1361         }
1362         jbd2_superblock_csum_set(journal, sb);
1363         get_bh(bh);
1364         bh->b_end_io = end_buffer_write_sync;
1365         ret = submit_bh(REQ_OP_WRITE, write_flags, bh);
1366         wait_on_buffer(bh);
1367         if (buffer_write_io_error(bh)) {
1368                 clear_buffer_write_io_error(bh);
1369                 set_buffer_uptodate(bh);
1370                 ret = -EIO;
1371         }
1372         if (ret) {
1373                 printk(KERN_ERR "JBD2: Error %d detected when updating "
1374                        "journal superblock for %s.\n", ret,
1375                        journal->j_devname);
1376                 jbd2_journal_abort(journal, ret);
1377         }
1378
1379         return ret;
1380 }
1381
1382 /**
1383  * jbd2_journal_update_sb_log_tail() - Update log tail in journal sb on disk.
1384  * @journal: The journal to update.
1385  * @tail_tid: TID of the new transaction at the tail of the log
1386  * @tail_block: The first block of the transaction at the tail of the log
1387  * @write_op: With which operation should we write the journal sb
1388  *
1389  * Update a journal's superblock information about log tail and write it to
1390  * disk, waiting for the IO to complete.
1391  */
1392 int jbd2_journal_update_sb_log_tail(journal_t *journal, tid_t tail_tid,
1393                                      unsigned long tail_block, int write_op)
1394 {
1395         journal_superblock_t *sb = journal->j_superblock;
1396         int ret;
1397
1398         BUG_ON(!mutex_is_locked(&journal->j_checkpoint_mutex));
1399         jbd_debug(1, "JBD2: updating superblock (start %lu, seq %u)\n",
1400                   tail_block, tail_tid);
1401
1402         sb->s_sequence = cpu_to_be32(tail_tid);
1403         sb->s_start    = cpu_to_be32(tail_block);
1404
1405         ret = jbd2_write_superblock(journal, write_op);
1406         if (ret)
1407                 goto out;
1408
1409         /* Log is no longer empty */
1410         write_lock(&journal->j_state_lock);
1411         WARN_ON(!sb->s_sequence);
1412         journal->j_flags &= ~JBD2_FLUSHED;
1413         write_unlock(&journal->j_state_lock);
1414
1415 out:
1416         return ret;
1417 }
1418
1419 /**
1420  * jbd2_mark_journal_empty() - Mark on disk journal as empty.
1421  * @journal: The journal to update.
1422  * @write_op: With which operation should we write the journal sb
1423  *
1424  * Update a journal's dynamic superblock fields to show that journal is empty.
1425  * Write updated superblock to disk waiting for IO to complete.
1426  */
1427 static void jbd2_mark_journal_empty(journal_t *journal, int write_op)
1428 {
1429         journal_superblock_t *sb = journal->j_superblock;
1430
1431         BUG_ON(!mutex_is_locked(&journal->j_checkpoint_mutex));
1432         read_lock(&journal->j_state_lock);
1433         /* Is it already empty? */
1434         if (sb->s_start == 0) {
1435                 read_unlock(&journal->j_state_lock);
1436                 return;
1437         }
1438         jbd_debug(1, "JBD2: Marking journal as empty (seq %d)\n",
1439                   journal->j_tail_sequence);
1440
1441         sb->s_sequence = cpu_to_be32(journal->j_tail_sequence);
1442         sb->s_start    = cpu_to_be32(0);
1443         read_unlock(&journal->j_state_lock);
1444
1445         jbd2_write_superblock(journal, write_op);
1446
1447         /* Log is no longer empty */
1448         write_lock(&journal->j_state_lock);
1449         journal->j_flags |= JBD2_FLUSHED;
1450         write_unlock(&journal->j_state_lock);
1451 }
1452
1453
1454 /**
1455  * jbd2_journal_update_sb_errno() - Update error in the journal.
1456  * @journal: The journal to update.
1457  *
1458  * Update a journal's errno.  Write updated superblock to disk waiting for IO
1459  * to complete.
1460  */
1461 void jbd2_journal_update_sb_errno(journal_t *journal)
1462 {
1463         journal_superblock_t *sb = journal->j_superblock;
1464
1465         read_lock(&journal->j_state_lock);
1466         jbd_debug(1, "JBD2: updating superblock error (errno %d)\n",
1467                   journal->j_errno);
1468         sb->s_errno    = cpu_to_be32(journal->j_errno);
1469         read_unlock(&journal->j_state_lock);
1470
1471         jbd2_write_superblock(journal, REQ_SYNC | REQ_FUA);
1472 }
1473 EXPORT_SYMBOL(jbd2_journal_update_sb_errno);
1474
1475 /*
1476  * Read the superblock for a given journal, performing initial
1477  * validation of the format.
1478  */
1479 static int journal_get_superblock(journal_t *journal)
1480 {
1481         struct buffer_head *bh;
1482         journal_superblock_t *sb;
1483         int err = -EIO;
1484
1485         bh = journal->j_sb_buffer;
1486
1487         J_ASSERT(bh != NULL);
1488         if (!buffer_uptodate(bh)) {
1489                 ll_rw_block(REQ_OP_READ, 0, 1, &bh);
1490                 wait_on_buffer(bh);
1491                 if (!buffer_uptodate(bh)) {
1492                         printk(KERN_ERR
1493                                 "JBD2: IO error reading journal superblock\n");
1494                         goto out;
1495                 }
1496         }
1497
1498         if (buffer_verified(bh))
1499                 return 0;
1500
1501         sb = journal->j_superblock;
1502
1503         err = -EINVAL;
1504
1505         if (sb->s_header.h_magic != cpu_to_be32(JBD2_MAGIC_NUMBER) ||
1506             sb->s_blocksize != cpu_to_be32(journal->j_blocksize)) {
1507                 printk(KERN_WARNING "JBD2: no valid journal superblock found\n");
1508                 goto out;
1509         }
1510
1511         switch(be32_to_cpu(sb->s_header.h_blocktype)) {
1512         case JBD2_SUPERBLOCK_V1:
1513                 journal->j_format_version = 1;
1514                 break;
1515         case JBD2_SUPERBLOCK_V2:
1516                 journal->j_format_version = 2;
1517                 break;
1518         default:
1519                 printk(KERN_WARNING "JBD2: unrecognised superblock format ID\n");
1520                 goto out;
1521         }
1522
1523         if (be32_to_cpu(sb->s_maxlen) < journal->j_maxlen)
1524                 journal->j_maxlen = be32_to_cpu(sb->s_maxlen);
1525         else if (be32_to_cpu(sb->s_maxlen) > journal->j_maxlen) {
1526                 printk(KERN_WARNING "JBD2: journal file too short\n");
1527                 goto out;
1528         }
1529
1530         if (be32_to_cpu(sb->s_first) == 0 ||
1531             be32_to_cpu(sb->s_first) >= journal->j_maxlen) {
1532                 printk(KERN_WARNING
1533                         "JBD2: Invalid start block of journal: %u\n",
1534                         be32_to_cpu(sb->s_first));
1535                 goto out;
1536         }
1537
1538         if (jbd2_has_feature_csum2(journal) &&
1539             jbd2_has_feature_csum3(journal)) {
1540                 /* Can't have checksum v2 and v3 at the same time! */
1541                 printk(KERN_ERR "JBD2: Can't enable checksumming v2 and v3 "
1542                        "at the same time!\n");
1543                 goto out;
1544         }
1545
1546         if (jbd2_journal_has_csum_v2or3_feature(journal) &&
1547             jbd2_has_feature_checksum(journal)) {
1548                 /* Can't have checksum v1 and v2 on at the same time! */
1549                 printk(KERN_ERR "JBD2: Can't enable checksumming v1 and v2/3 "
1550                        "at the same time!\n");
1551                 goto out;
1552         }
1553
1554         if (!jbd2_verify_csum_type(journal, sb)) {
1555                 printk(KERN_ERR "JBD2: Unknown checksum type\n");
1556                 goto out;
1557         }
1558
1559         /* Load the checksum driver */
1560         if (jbd2_journal_has_csum_v2or3_feature(journal)) {
1561                 journal->j_chksum_driver = crypto_alloc_shash("crc32c", 0, 0);
1562                 if (IS_ERR(journal->j_chksum_driver)) {
1563                         printk(KERN_ERR "JBD2: Cannot load crc32c driver.\n");
1564                         err = PTR_ERR(journal->j_chksum_driver);
1565                         journal->j_chksum_driver = NULL;
1566                         goto out;
1567                 }
1568         }
1569
1570         /* Check superblock checksum */
1571         if (!jbd2_superblock_csum_verify(journal, sb)) {
1572                 printk(KERN_ERR "JBD2: journal checksum error\n");
1573                 err = -EFSBADCRC;
1574                 goto out;
1575         }
1576
1577         /* Precompute checksum seed for all metadata */
1578         if (jbd2_journal_has_csum_v2or3(journal))
1579                 journal->j_csum_seed = jbd2_chksum(journal, ~0, sb->s_uuid,
1580                                                    sizeof(sb->s_uuid));
1581
1582         set_buffer_verified(bh);
1583
1584         return 0;
1585
1586 out:
1587         journal_fail_superblock(journal);
1588         return err;
1589 }
1590
1591 /*
1592  * Load the on-disk journal superblock and read the key fields into the
1593  * journal_t.
1594  */
1595
1596 static int load_superblock(journal_t *journal)
1597 {
1598         int err;
1599         journal_superblock_t *sb;
1600
1601         err = journal_get_superblock(journal);
1602         if (err)
1603                 return err;
1604
1605         sb = journal->j_superblock;
1606
1607         journal->j_tail_sequence = be32_to_cpu(sb->s_sequence);
1608         journal->j_tail = be32_to_cpu(sb->s_start);
1609         journal->j_first = be32_to_cpu(sb->s_first);
1610         journal->j_last = be32_to_cpu(sb->s_maxlen);
1611         journal->j_errno = be32_to_cpu(sb->s_errno);
1612
1613         return 0;
1614 }
1615
1616
1617 /**
1618  * int jbd2_journal_load() - Read journal from disk.
1619  * @journal: Journal to act on.
1620  *
1621  * Given a journal_t structure which tells us which disk blocks contain
1622  * a journal, read the journal from disk to initialise the in-memory
1623  * structures.
1624  */
1625 int jbd2_journal_load(journal_t *journal)
1626 {
1627         int err;
1628         journal_superblock_t *sb;
1629
1630         err = load_superblock(journal);
1631         if (err)
1632                 return err;
1633
1634         sb = journal->j_superblock;
1635         /* If this is a V2 superblock, then we have to check the
1636          * features flags on it. */
1637
1638         if (journal->j_format_version >= 2) {
1639                 if ((sb->s_feature_ro_compat &
1640                      ~cpu_to_be32(JBD2_KNOWN_ROCOMPAT_FEATURES)) ||
1641                     (sb->s_feature_incompat &
1642                      ~cpu_to_be32(JBD2_KNOWN_INCOMPAT_FEATURES))) {
1643                         printk(KERN_WARNING
1644                                 "JBD2: Unrecognised features on journal\n");
1645                         return -EINVAL;
1646                 }
1647         }
1648
1649         /*
1650          * Create a slab for this blocksize
1651          */
1652         err = jbd2_journal_create_slab(be32_to_cpu(sb->s_blocksize));
1653         if (err)
1654                 return err;
1655
1656         /* Let the recovery code check whether it needs to recover any
1657          * data from the journal. */
1658         if (jbd2_journal_recover(journal))
1659                 goto recovery_error;
1660
1661         if (journal->j_failed_commit) {
1662                 printk(KERN_ERR "JBD2: journal transaction %u on %s "
1663                        "is corrupt.\n", journal->j_failed_commit,
1664                        journal->j_devname);
1665                 return -EFSCORRUPTED;
1666         }
1667
1668         /* OK, we've finished with the dynamic journal bits:
1669          * reinitialise the dynamic contents of the superblock in memory
1670          * and reset them on disk. */
1671         if (journal_reset(journal))
1672                 goto recovery_error;
1673
1674         journal->j_flags &= ~JBD2_ABORT;
1675         journal->j_flags |= JBD2_LOADED;
1676         return 0;
1677
1678 recovery_error:
1679         printk(KERN_WARNING "JBD2: recovery failed\n");
1680         return -EIO;
1681 }
1682
1683 /**
1684  * void jbd2_journal_destroy() - Release a journal_t structure.
1685  * @journal: Journal to act on.
1686  *
1687  * Release a journal_t structure once it is no longer in use by the
1688  * journaled object.
1689  * Return <0 if we couldn't clean up the journal.
1690  */
1691 int jbd2_journal_destroy(journal_t *journal)
1692 {
1693         int err = 0;
1694
1695         /* Wait for the commit thread to wake up and die. */
1696         journal_kill_thread(journal);
1697
1698         /* Force a final log commit */
1699         if (journal->j_running_transaction)
1700                 jbd2_journal_commit_transaction(journal);
1701
1702         /* Force any old transactions to disk */
1703
1704         /* Totally anal locking here... */
1705         spin_lock(&journal->j_list_lock);
1706         while (journal->j_checkpoint_transactions != NULL) {
1707                 spin_unlock(&journal->j_list_lock);
1708                 mutex_lock_io(&journal->j_checkpoint_mutex);
1709                 err = jbd2_log_do_checkpoint(journal);
1710                 mutex_unlock(&journal->j_checkpoint_mutex);
1711                 /*
1712                  * If checkpointing failed, just free the buffers to avoid
1713                  * looping forever
1714                  */
1715                 if (err) {
1716                         jbd2_journal_destroy_checkpoint(journal);
1717                         spin_lock(&journal->j_list_lock);
1718                         break;
1719                 }
1720                 spin_lock(&journal->j_list_lock);
1721         }
1722
1723         J_ASSERT(journal->j_running_transaction == NULL);
1724         J_ASSERT(journal->j_committing_transaction == NULL);
1725         J_ASSERT(journal->j_checkpoint_transactions == NULL);
1726         spin_unlock(&journal->j_list_lock);
1727
1728         if (journal->j_sb_buffer) {
1729                 if (!is_journal_aborted(journal)) {
1730                         mutex_lock_io(&journal->j_checkpoint_mutex);
1731
1732                         write_lock(&journal->j_state_lock);
1733                         journal->j_tail_sequence =
1734                                 ++journal->j_transaction_sequence;
1735                         write_unlock(&journal->j_state_lock);
1736
1737                         jbd2_mark_journal_empty(journal,
1738                                         REQ_SYNC | REQ_PREFLUSH | REQ_FUA);
1739                         mutex_unlock(&journal->j_checkpoint_mutex);
1740                 } else
1741                         err = -EIO;
1742                 brelse(journal->j_sb_buffer);
1743         }
1744
1745         if (journal->j_proc_entry)
1746                 jbd2_stats_proc_exit(journal);
1747         iput(journal->j_inode);
1748         if (journal->j_revoke)
1749                 jbd2_journal_destroy_revoke(journal);
1750         if (journal->j_chksum_driver)
1751                 crypto_free_shash(journal->j_chksum_driver);
1752         kfree(journal->j_wbuf);
1753         kfree(journal);
1754
1755         return err;
1756 }
1757
1758
1759 /**
1760  *int jbd2_journal_check_used_features () - Check if features specified are used.
1761  * @journal: Journal to check.
1762  * @compat: bitmask of compatible features
1763  * @ro: bitmask of features that force read-only mount
1764  * @incompat: bitmask of incompatible features
1765  *
1766  * Check whether the journal uses all of a given set of
1767  * features.  Return true (non-zero) if it does.
1768  **/
1769
1770 int jbd2_journal_check_used_features (journal_t *journal, unsigned long compat,
1771                                  unsigned long ro, unsigned long incompat)
1772 {
1773         journal_superblock_t *sb;
1774
1775         if (!compat && !ro && !incompat)
1776                 return 1;
1777         /* Load journal superblock if it is not loaded yet. */
1778         if (journal->j_format_version == 0 &&
1779             journal_get_superblock(journal) != 0)
1780                 return 0;
1781         if (journal->j_format_version == 1)
1782                 return 0;
1783
1784         sb = journal->j_superblock;
1785
1786         if (((be32_to_cpu(sb->s_feature_compat) & compat) == compat) &&
1787             ((be32_to_cpu(sb->s_feature_ro_compat) & ro) == ro) &&
1788             ((be32_to_cpu(sb->s_feature_incompat) & incompat) == incompat))
1789                 return 1;
1790
1791         return 0;
1792 }
1793
1794 /**
1795  * int jbd2_journal_check_available_features() - Check feature set in journalling layer
1796  * @journal: Journal to check.
1797  * @compat: bitmask of compatible features
1798  * @ro: bitmask of features that force read-only mount
1799  * @incompat: bitmask of incompatible features
1800  *
1801  * Check whether the journaling code supports the use of
1802  * all of a given set of features on this journal.  Return true
1803  * (non-zero) if it can. */
1804
1805 int jbd2_journal_check_available_features (journal_t *journal, unsigned long compat,
1806                                       unsigned long ro, unsigned long incompat)
1807 {
1808         if (!compat && !ro && !incompat)
1809                 return 1;
1810
1811         /* We can support any known requested features iff the
1812          * superblock is in version 2.  Otherwise we fail to support any
1813          * extended sb features. */
1814
1815         if (journal->j_format_version != 2)
1816                 return 0;
1817
1818         if ((compat   & JBD2_KNOWN_COMPAT_FEATURES) == compat &&
1819             (ro       & JBD2_KNOWN_ROCOMPAT_FEATURES) == ro &&
1820             (incompat & JBD2_KNOWN_INCOMPAT_FEATURES) == incompat)
1821                 return 1;
1822
1823         return 0;
1824 }
1825
1826 /**
1827  * int jbd2_journal_set_features () - Mark a given journal feature in the superblock
1828  * @journal: Journal to act on.
1829  * @compat: bitmask of compatible features
1830  * @ro: bitmask of features that force read-only mount
1831  * @incompat: bitmask of incompatible features
1832  *
1833  * Mark a given journal feature as present on the
1834  * superblock.  Returns true if the requested features could be set.
1835  *
1836  */
1837
1838 int jbd2_journal_set_features (journal_t *journal, unsigned long compat,
1839                           unsigned long ro, unsigned long incompat)
1840 {
1841 #define INCOMPAT_FEATURE_ON(f) \
1842                 ((incompat & (f)) && !(sb->s_feature_incompat & cpu_to_be32(f)))
1843 #define COMPAT_FEATURE_ON(f) \
1844                 ((compat & (f)) && !(sb->s_feature_compat & cpu_to_be32(f)))
1845         journal_superblock_t *sb;
1846
1847         if (jbd2_journal_check_used_features(journal, compat, ro, incompat))
1848                 return 1;
1849
1850         if (!jbd2_journal_check_available_features(journal, compat, ro, incompat))
1851                 return 0;
1852
1853         /* If enabling v2 checksums, turn on v3 instead */
1854         if (incompat & JBD2_FEATURE_INCOMPAT_CSUM_V2) {
1855                 incompat &= ~JBD2_FEATURE_INCOMPAT_CSUM_V2;
1856                 incompat |= JBD2_FEATURE_INCOMPAT_CSUM_V3;
1857         }
1858
1859         /* Asking for checksumming v3 and v1?  Only give them v3. */
1860         if (incompat & JBD2_FEATURE_INCOMPAT_CSUM_V3 &&
1861             compat & JBD2_FEATURE_COMPAT_CHECKSUM)
1862                 compat &= ~JBD2_FEATURE_COMPAT_CHECKSUM;
1863
1864         jbd_debug(1, "Setting new features 0x%lx/0x%lx/0x%lx\n",
1865                   compat, ro, incompat);
1866
1867         sb = journal->j_superblock;
1868
1869         /* If enabling v3 checksums, update superblock */
1870         if (INCOMPAT_FEATURE_ON(JBD2_FEATURE_INCOMPAT_CSUM_V3)) {
1871                 sb->s_checksum_type = JBD2_CRC32C_CHKSUM;
1872                 sb->s_feature_compat &=
1873                         ~cpu_to_be32(JBD2_FEATURE_COMPAT_CHECKSUM);
1874
1875                 /* Load the checksum driver */
1876                 if (journal->j_chksum_driver == NULL) {
1877                         journal->j_chksum_driver = crypto_alloc_shash("crc32c",
1878                                                                       0, 0);
1879                         if (IS_ERR(journal->j_chksum_driver)) {
1880                                 printk(KERN_ERR "JBD2: Cannot load crc32c "
1881                                        "driver.\n");
1882                                 journal->j_chksum_driver = NULL;
1883                                 return 0;
1884                         }
1885
1886                         /* Precompute checksum seed for all metadata */
1887                         journal->j_csum_seed = jbd2_chksum(journal, ~0,
1888                                                            sb->s_uuid,
1889                                                            sizeof(sb->s_uuid));
1890                 }
1891         }
1892
1893         /* If enabling v1 checksums, downgrade superblock */
1894         if (COMPAT_FEATURE_ON(JBD2_FEATURE_COMPAT_CHECKSUM))
1895                 sb->s_feature_incompat &=
1896                         ~cpu_to_be32(JBD2_FEATURE_INCOMPAT_CSUM_V2 |
1897                                      JBD2_FEATURE_INCOMPAT_CSUM_V3);
1898
1899         sb->s_feature_compat    |= cpu_to_be32(compat);
1900         sb->s_feature_ro_compat |= cpu_to_be32(ro);
1901         sb->s_feature_incompat  |= cpu_to_be32(incompat);
1902
1903         return 1;
1904 #undef COMPAT_FEATURE_ON
1905 #undef INCOMPAT_FEATURE_ON
1906 }
1907
1908 /*
1909  * jbd2_journal_clear_features () - Clear a given journal feature in the
1910  *                                  superblock
1911  * @journal: Journal to act on.
1912  * @compat: bitmask of compatible features
1913  * @ro: bitmask of features that force read-only mount
1914  * @incompat: bitmask of incompatible features
1915  *
1916  * Clear a given journal feature as present on the
1917  * superblock.
1918  */
1919 void jbd2_journal_clear_features(journal_t *journal, unsigned long compat,
1920                                 unsigned long ro, unsigned long incompat)
1921 {
1922         journal_superblock_t *sb;
1923
1924         jbd_debug(1, "Clear features 0x%lx/0x%lx/0x%lx\n",
1925                   compat, ro, incompat);
1926
1927         sb = journal->j_superblock;
1928
1929         sb->s_feature_compat    &= ~cpu_to_be32(compat);
1930         sb->s_feature_ro_compat &= ~cpu_to_be32(ro);
1931         sb->s_feature_incompat  &= ~cpu_to_be32(incompat);
1932 }
1933 EXPORT_SYMBOL(jbd2_journal_clear_features);
1934
1935 /**
1936  * int jbd2_journal_flush () - Flush journal
1937  * @journal: Journal to act on.
1938  *
1939  * Flush all data for a given journal to disk and empty the journal.
1940  * Filesystems can use this when remounting readonly to ensure that
1941  * recovery does not need to happen on remount.
1942  */
1943
1944 int jbd2_journal_flush(journal_t *journal)
1945 {
1946         int err = 0;
1947         transaction_t *transaction = NULL;
1948
1949         write_lock(&journal->j_state_lock);
1950
1951         /* Force everything buffered to the log... */
1952         if (journal->j_running_transaction) {
1953                 transaction = journal->j_running_transaction;
1954                 __jbd2_log_start_commit(journal, transaction->t_tid);
1955         } else if (journal->j_committing_transaction)
1956                 transaction = journal->j_committing_transaction;
1957
1958         /* Wait for the log commit to complete... */
1959         if (transaction) {
1960                 tid_t tid = transaction->t_tid;
1961
1962                 write_unlock(&journal->j_state_lock);
1963                 jbd2_log_wait_commit(journal, tid);
1964         } else {
1965                 write_unlock(&journal->j_state_lock);
1966         }
1967
1968         /* ...and flush everything in the log out to disk. */
1969         spin_lock(&journal->j_list_lock);
1970         while (!err && journal->j_checkpoint_transactions != NULL) {
1971                 spin_unlock(&journal->j_list_lock);
1972                 mutex_lock_io(&journal->j_checkpoint_mutex);
1973                 err = jbd2_log_do_checkpoint(journal);
1974                 mutex_unlock(&journal->j_checkpoint_mutex);
1975                 spin_lock(&journal->j_list_lock);
1976         }
1977         spin_unlock(&journal->j_list_lock);
1978
1979         if (is_journal_aborted(journal))
1980                 return -EIO;
1981
1982         mutex_lock_io(&journal->j_checkpoint_mutex);
1983         if (!err) {
1984                 err = jbd2_cleanup_journal_tail(journal);
1985                 if (err < 0) {
1986                         mutex_unlock(&journal->j_checkpoint_mutex);
1987                         goto out;
1988                 }
1989                 err = 0;
1990         }
1991
1992         /* Finally, mark the journal as really needing no recovery.
1993          * This sets s_start==0 in the underlying superblock, which is
1994          * the magic code for a fully-recovered superblock.  Any future
1995          * commits of data to the journal will restore the current
1996          * s_start value. */
1997         jbd2_mark_journal_empty(journal, REQ_SYNC | REQ_FUA);
1998         mutex_unlock(&journal->j_checkpoint_mutex);
1999         write_lock(&journal->j_state_lock);
2000         J_ASSERT(!journal->j_running_transaction);
2001         J_ASSERT(!journal->j_committing_transaction);
2002         J_ASSERT(!journal->j_checkpoint_transactions);
2003         J_ASSERT(journal->j_head == journal->j_tail);
2004         J_ASSERT(journal->j_tail_sequence == journal->j_transaction_sequence);
2005         write_unlock(&journal->j_state_lock);
2006 out:
2007         return err;
2008 }
2009
2010 /**
2011  * int jbd2_journal_wipe() - Wipe journal contents
2012  * @journal: Journal to act on.
2013  * @write: flag (see below)
2014  *
2015  * Wipe out all of the contents of a journal, safely.  This will produce
2016  * a warning if the journal contains any valid recovery information.
2017  * Must be called between journal_init_*() and jbd2_journal_load().
2018  *
2019  * If 'write' is non-zero, then we wipe out the journal on disk; otherwise
2020  * we merely suppress recovery.
2021  */
2022
2023 int jbd2_journal_wipe(journal_t *journal, int write)
2024 {
2025         int err = 0;
2026
2027         J_ASSERT (!(journal->j_flags & JBD2_LOADED));
2028
2029         err = load_superblock(journal);
2030         if (err)
2031                 return err;
2032
2033         if (!journal->j_tail)
2034                 goto no_recovery;
2035
2036         printk(KERN_WARNING "JBD2: %s recovery information on journal\n",
2037                 write ? "Clearing" : "Ignoring");
2038
2039         err = jbd2_journal_skip_recovery(journal);
2040         if (write) {
2041                 /* Lock to make assertions happy... */
2042                 mutex_lock(&journal->j_checkpoint_mutex);
2043                 jbd2_mark_journal_empty(journal, REQ_SYNC | REQ_FUA);
2044                 mutex_unlock(&journal->j_checkpoint_mutex);
2045         }
2046
2047  no_recovery:
2048         return err;
2049 }
2050
2051 /*
2052  * Journal abort has very specific semantics, which we describe
2053  * for journal abort.
2054  *
2055  * Two internal functions, which provide abort to the jbd layer
2056  * itself are here.
2057  */
2058
2059 /*
2060  * Quick version for internal journal use (doesn't lock the journal).
2061  * Aborts hard --- we mark the abort as occurred, but do _nothing_ else,
2062  * and don't attempt to make any other journal updates.
2063  */
2064 void __jbd2_journal_abort_hard(journal_t *journal)
2065 {
2066         transaction_t *transaction;
2067
2068         if (journal->j_flags & JBD2_ABORT)
2069                 return;
2070
2071         printk(KERN_ERR "Aborting journal on device %s.\n",
2072                journal->j_devname);
2073
2074         write_lock(&journal->j_state_lock);
2075         journal->j_flags |= JBD2_ABORT;
2076         transaction = journal->j_running_transaction;
2077         if (transaction)
2078                 __jbd2_log_start_commit(journal, transaction->t_tid);
2079         write_unlock(&journal->j_state_lock);
2080 }
2081
2082 /* Soft abort: record the abort error status in the journal superblock,
2083  * but don't do any other IO. */
2084 static void __journal_abort_soft (journal_t *journal, int errno)
2085 {
2086         if (journal->j_flags & JBD2_ABORT)
2087                 return;
2088
2089         if (!journal->j_errno)
2090                 journal->j_errno = errno;
2091
2092         __jbd2_journal_abort_hard(journal);
2093
2094         if (errno) {
2095                 jbd2_journal_update_sb_errno(journal);
2096                 write_lock(&journal->j_state_lock);
2097                 journal->j_flags |= JBD2_REC_ERR;
2098                 write_unlock(&journal->j_state_lock);
2099         }
2100 }
2101
2102 /**
2103  * void jbd2_journal_abort () - Shutdown the journal immediately.
2104  * @journal: the journal to shutdown.
2105  * @errno:   an error number to record in the journal indicating
2106  *           the reason for the shutdown.
2107  *
2108  * Perform a complete, immediate shutdown of the ENTIRE
2109  * journal (not of a single transaction).  This operation cannot be
2110  * undone without closing and reopening the journal.
2111  *
2112  * The jbd2_journal_abort function is intended to support higher level error
2113  * recovery mechanisms such as the ext2/ext3 remount-readonly error
2114  * mode.
2115  *
2116  * Journal abort has very specific semantics.  Any existing dirty,
2117  * unjournaled buffers in the main filesystem will still be written to
2118  * disk by bdflush, but the journaling mechanism will be suspended
2119  * immediately and no further transaction commits will be honoured.
2120  *
2121  * Any dirty, journaled buffers will be written back to disk without
2122  * hitting the journal.  Atomicity cannot be guaranteed on an aborted
2123  * filesystem, but we _do_ attempt to leave as much data as possible
2124  * behind for fsck to use for cleanup.
2125  *
2126  * Any attempt to get a new transaction handle on a journal which is in
2127  * ABORT state will just result in an -EROFS error return.  A
2128  * jbd2_journal_stop on an existing handle will return -EIO if we have
2129  * entered abort state during the update.
2130  *
2131  * Recursive transactions are not disturbed by journal abort until the
2132  * final jbd2_journal_stop, which will receive the -EIO error.
2133  *
2134  * Finally, the jbd2_journal_abort call allows the caller to supply an errno
2135  * which will be recorded (if possible) in the journal superblock.  This
2136  * allows a client to record failure conditions in the middle of a
2137  * transaction without having to complete the transaction to record the
2138  * failure to disk.  ext3_error, for example, now uses this
2139  * functionality.
2140  *
2141  * Errors which originate from within the journaling layer will NOT
2142  * supply an errno; a null errno implies that absolutely no further
2143  * writes are done to the journal (unless there are any already in
2144  * progress).
2145  *
2146  */
2147
2148 void jbd2_journal_abort(journal_t *journal, int errno)
2149 {
2150         __journal_abort_soft(journal, errno);
2151 }
2152
2153 /**
2154  * int jbd2_journal_errno () - returns the journal's error state.
2155  * @journal: journal to examine.
2156  *
2157  * This is the errno number set with jbd2_journal_abort(), the last
2158  * time the journal was mounted - if the journal was stopped
2159  * without calling abort this will be 0.
2160  *
2161  * If the journal has been aborted on this mount time -EROFS will
2162  * be returned.
2163  */
2164 int jbd2_journal_errno(journal_t *journal)
2165 {
2166         int err;
2167
2168         read_lock(&journal->j_state_lock);
2169         if (journal->j_flags & JBD2_ABORT)
2170                 err = -EROFS;
2171         else
2172                 err = journal->j_errno;
2173         read_unlock(&journal->j_state_lock);
2174         return err;
2175 }
2176
2177 /**
2178  * int jbd2_journal_clear_err () - clears the journal's error state
2179  * @journal: journal to act on.
2180  *
2181  * An error must be cleared or acked to take a FS out of readonly
2182  * mode.
2183  */
2184 int jbd2_journal_clear_err(journal_t *journal)
2185 {
2186         int err = 0;
2187
2188         write_lock(&journal->j_state_lock);
2189         if (journal->j_flags & JBD2_ABORT)
2190                 err = -EROFS;
2191         else
2192                 journal->j_errno = 0;
2193         write_unlock(&journal->j_state_lock);
2194         return err;
2195 }
2196
2197 /**
2198  * void jbd2_journal_ack_err() - Ack journal err.
2199  * @journal: journal to act on.
2200  *
2201  * An error must be cleared or acked to take a FS out of readonly
2202  * mode.
2203  */
2204 void jbd2_journal_ack_err(journal_t *journal)
2205 {
2206         write_lock(&journal->j_state_lock);
2207         if (journal->j_errno)
2208                 journal->j_flags |= JBD2_ACK_ERR;
2209         write_unlock(&journal->j_state_lock);
2210 }
2211
2212 int jbd2_journal_blocks_per_page(struct inode *inode)
2213 {
2214         return 1 << (PAGE_SHIFT - inode->i_sb->s_blocksize_bits);
2215 }
2216
2217 /*
2218  * helper functions to deal with 32 or 64bit block numbers.
2219  */
2220 size_t journal_tag_bytes(journal_t *journal)
2221 {
2222         size_t sz;
2223
2224         if (jbd2_has_feature_csum3(journal))
2225                 return sizeof(journal_block_tag3_t);
2226
2227         sz = sizeof(journal_block_tag_t);
2228
2229         if (jbd2_has_feature_csum2(journal))
2230                 sz += sizeof(__u16);
2231
2232         if (jbd2_has_feature_64bit(journal))
2233                 return sz;
2234         else
2235                 return sz - sizeof(__u32);
2236 }
2237
2238 /*
2239  * JBD memory management
2240  *
2241  * These functions are used to allocate block-sized chunks of memory
2242  * used for making copies of buffer_head data.  Very often it will be
2243  * page-sized chunks of data, but sometimes it will be in
2244  * sub-page-size chunks.  (For example, 16k pages on Power systems
2245  * with a 4k block file system.)  For blocks smaller than a page, we
2246  * use a SLAB allocator.  There are slab caches for each block size,
2247  * which are allocated at mount time, if necessary, and we only free
2248  * (all of) the slab caches when/if the jbd2 module is unloaded.  For
2249  * this reason we don't need to a mutex to protect access to
2250  * jbd2_slab[] allocating or releasing memory; only in
2251  * jbd2_journal_create_slab().
2252  */
2253 #define JBD2_MAX_SLABS 8
2254 static struct kmem_cache *jbd2_slab[JBD2_MAX_SLABS];
2255
2256 static const char *jbd2_slab_names[JBD2_MAX_SLABS] = {
2257         "jbd2_1k", "jbd2_2k", "jbd2_4k", "jbd2_8k",
2258         "jbd2_16k", "jbd2_32k", "jbd2_64k", "jbd2_128k"
2259 };
2260
2261
2262 static void jbd2_journal_destroy_slabs(void)
2263 {
2264         int i;
2265
2266         for (i = 0; i < JBD2_MAX_SLABS; i++) {
2267                 if (jbd2_slab[i])
2268                         kmem_cache_destroy(jbd2_slab[i]);
2269                 jbd2_slab[i] = NULL;
2270         }
2271 }
2272
2273 static int jbd2_journal_create_slab(size_t size)
2274 {
2275         static DEFINE_MUTEX(jbd2_slab_create_mutex);
2276         int i = order_base_2(size) - 10;
2277         size_t slab_size;
2278
2279         if (size == PAGE_SIZE)
2280                 return 0;
2281
2282         if (i >= JBD2_MAX_SLABS)
2283                 return -EINVAL;
2284
2285         if (unlikely(i < 0))
2286                 i = 0;
2287         mutex_lock(&jbd2_slab_create_mutex);
2288         if (jbd2_slab[i]) {
2289                 mutex_unlock(&jbd2_slab_create_mutex);
2290                 return 0;       /* Already created */
2291         }
2292
2293         slab_size = 1 << (i+10);
2294         jbd2_slab[i] = kmem_cache_create(jbd2_slab_names[i], slab_size,
2295                                          slab_size, 0, NULL);
2296         mutex_unlock(&jbd2_slab_create_mutex);
2297         if (!jbd2_slab[i]) {
2298                 printk(KERN_EMERG "JBD2: no memory for jbd2_slab cache\n");
2299                 return -ENOMEM;
2300         }
2301         return 0;
2302 }
2303
2304 static struct kmem_cache *get_slab(size_t size)
2305 {
2306         int i = order_base_2(size) - 10;
2307
2308         BUG_ON(i >= JBD2_MAX_SLABS);
2309         if (unlikely(i < 0))
2310                 i = 0;
2311         BUG_ON(jbd2_slab[i] == NULL);
2312         return jbd2_slab[i];
2313 }
2314
2315 void *jbd2_alloc(size_t size, gfp_t flags)
2316 {
2317         void *ptr;
2318
2319         BUG_ON(size & (size-1)); /* Must be a power of 2 */
2320
2321         if (size < PAGE_SIZE)
2322                 ptr = kmem_cache_alloc(get_slab(size), flags);
2323         else
2324                 ptr = (void *)__get_free_pages(flags, get_order(size));
2325
2326         /* Check alignment; SLUB has gotten this wrong in the past,
2327          * and this can lead to user data corruption! */
2328         BUG_ON(((unsigned long) ptr) & (size-1));
2329
2330         return ptr;
2331 }
2332
2333 void jbd2_free(void *ptr, size_t size)
2334 {
2335         if (size < PAGE_SIZE)
2336                 kmem_cache_free(get_slab(size), ptr);
2337         else
2338                 free_pages((unsigned long)ptr, get_order(size));
2339 };
2340
2341 /*
2342  * Journal_head storage management
2343  */
2344 static struct kmem_cache *jbd2_journal_head_cache;
2345 #ifdef CONFIG_JBD2_DEBUG
2346 static atomic_t nr_journal_heads = ATOMIC_INIT(0);
2347 #endif
2348
2349 static int jbd2_journal_init_journal_head_cache(void)
2350 {
2351         int retval;
2352
2353         J_ASSERT(jbd2_journal_head_cache == NULL);
2354         jbd2_journal_head_cache = kmem_cache_create("jbd2_journal_head",
2355                                 sizeof(struct journal_head),
2356                                 0,              /* offset */
2357                                 SLAB_TEMPORARY | SLAB_DESTROY_BY_RCU,
2358                                 NULL);          /* ctor */
2359         retval = 0;
2360         if (!jbd2_journal_head_cache) {
2361                 retval = -ENOMEM;
2362                 printk(KERN_EMERG "JBD2: no memory for journal_head cache\n");
2363         }
2364         return retval;
2365 }
2366
2367 static void jbd2_journal_destroy_journal_head_cache(void)
2368 {
2369         if (jbd2_journal_head_cache) {
2370                 kmem_cache_destroy(jbd2_journal_head_cache);
2371                 jbd2_journal_head_cache = NULL;
2372         }
2373 }
2374
2375 /*
2376  * journal_head splicing and dicing
2377  */
2378 static struct journal_head *journal_alloc_journal_head(void)
2379 {
2380         struct journal_head *ret;
2381
2382 #ifdef CONFIG_JBD2_DEBUG
2383         atomic_inc(&nr_journal_heads);
2384 #endif
2385         ret = kmem_cache_zalloc(jbd2_journal_head_cache, GFP_NOFS);
2386         if (!ret) {
2387                 jbd_debug(1, "out of memory for journal_head\n");
2388                 pr_notice_ratelimited("ENOMEM in %s, retrying.\n", __func__);
2389                 ret = kmem_cache_zalloc(jbd2_journal_head_cache,
2390                                 GFP_NOFS | __GFP_NOFAIL);
2391         }
2392         return ret;
2393 }
2394
2395 static void journal_free_journal_head(struct journal_head *jh)
2396 {
2397 #ifdef CONFIG_JBD2_DEBUG
2398         atomic_dec(&nr_journal_heads);
2399         memset(jh, JBD2_POISON_FREE, sizeof(*jh));
2400 #endif
2401         kmem_cache_free(jbd2_journal_head_cache, jh);
2402 }
2403
2404 /*
2405  * A journal_head is attached to a buffer_head whenever JBD has an
2406  * interest in the buffer.
2407  *
2408  * Whenever a buffer has an attached journal_head, its ->b_state:BH_JBD bit
2409  * is set.  This bit is tested in core kernel code where we need to take
2410  * JBD-specific actions.  Testing the zeroness of ->b_private is not reliable
2411  * there.
2412  *
2413  * When a buffer has its BH_JBD bit set, its ->b_count is elevated by one.
2414  *
2415  * When a buffer has its BH_JBD bit set it is immune from being released by
2416  * core kernel code, mainly via ->b_count.
2417  *
2418  * A journal_head is detached from its buffer_head when the journal_head's
2419  * b_jcount reaches zero. Running transaction (b_transaction) and checkpoint
2420  * transaction (b_cp_transaction) hold their references to b_jcount.
2421  *
2422  * Various places in the kernel want to attach a journal_head to a buffer_head
2423  * _before_ attaching the journal_head to a transaction.  To protect the
2424  * journal_head in this situation, jbd2_journal_add_journal_head elevates the
2425  * journal_head's b_jcount refcount by one.  The caller must call
2426  * jbd2_journal_put_journal_head() to undo this.
2427  *
2428  * So the typical usage would be:
2429  *
2430  *      (Attach a journal_head if needed.  Increments b_jcount)
2431  *      struct journal_head *jh = jbd2_journal_add_journal_head(bh);
2432  *      ...
2433  *      (Get another reference for transaction)
2434  *      jbd2_journal_grab_journal_head(bh);
2435  *      jh->b_transaction = xxx;
2436  *      (Put original reference)
2437  *      jbd2_journal_put_journal_head(jh);
2438  */
2439
2440 /*
2441  * Give a buffer_head a journal_head.
2442  *
2443  * May sleep.
2444  */
2445 struct journal_head *jbd2_journal_add_journal_head(struct buffer_head *bh)
2446 {
2447         struct journal_head *jh;
2448         struct journal_head *new_jh = NULL;
2449
2450 repeat:
2451         if (!buffer_jbd(bh))
2452                 new_jh = journal_alloc_journal_head();
2453
2454         jbd_lock_bh_journal_head(bh);
2455         if (buffer_jbd(bh)) {
2456                 jh = bh2jh(bh);
2457         } else {
2458                 J_ASSERT_BH(bh,
2459                         (atomic_read(&bh->b_count) > 0) ||
2460                         (bh->b_page && bh->b_page->mapping));
2461
2462                 if (!new_jh) {
2463                         jbd_unlock_bh_journal_head(bh);
2464                         goto repeat;
2465                 }
2466
2467                 jh = new_jh;
2468                 new_jh = NULL;          /* We consumed it */
2469                 set_buffer_jbd(bh);
2470                 bh->b_private = jh;
2471                 jh->b_bh = bh;
2472                 get_bh(bh);
2473                 BUFFER_TRACE(bh, "added journal_head");
2474         }
2475         jh->b_jcount++;
2476         jbd_unlock_bh_journal_head(bh);
2477         if (new_jh)
2478                 journal_free_journal_head(new_jh);
2479         return bh->b_private;
2480 }
2481
2482 /*
2483  * Grab a ref against this buffer_head's journal_head.  If it ended up not
2484  * having a journal_head, return NULL
2485  */
2486 struct journal_head *jbd2_journal_grab_journal_head(struct buffer_head *bh)
2487 {
2488         struct journal_head *jh = NULL;
2489
2490         jbd_lock_bh_journal_head(bh);
2491         if (buffer_jbd(bh)) {
2492                 jh = bh2jh(bh);
2493                 jh->b_jcount++;
2494         }
2495         jbd_unlock_bh_journal_head(bh);
2496         return jh;
2497 }
2498
2499 static void __journal_remove_journal_head(struct buffer_head *bh)
2500 {
2501         struct journal_head *jh = bh2jh(bh);
2502
2503         J_ASSERT_JH(jh, jh->b_jcount >= 0);
2504         J_ASSERT_JH(jh, jh->b_transaction == NULL);
2505         J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
2506         J_ASSERT_JH(jh, jh->b_cp_transaction == NULL);
2507         J_ASSERT_JH(jh, jh->b_jlist == BJ_None);
2508         J_ASSERT_BH(bh, buffer_jbd(bh));
2509         J_ASSERT_BH(bh, jh2bh(jh) == bh);
2510         BUFFER_TRACE(bh, "remove journal_head");
2511         if (jh->b_frozen_data) {
2512                 printk(KERN_WARNING "%s: freeing b_frozen_data\n", __func__);
2513                 jbd2_free(jh->b_frozen_data, bh->b_size);
2514         }
2515         if (jh->b_committed_data) {
2516                 printk(KERN_WARNING "%s: freeing b_committed_data\n", __func__);
2517                 jbd2_free(jh->b_committed_data, bh->b_size);
2518         }
2519         bh->b_private = NULL;
2520         jh->b_bh = NULL;        /* debug, really */
2521         clear_buffer_jbd(bh);
2522         journal_free_journal_head(jh);
2523 }
2524
2525 /*
2526  * Drop a reference on the passed journal_head.  If it fell to zero then
2527  * release the journal_head from the buffer_head.
2528  */
2529 void jbd2_journal_put_journal_head(struct journal_head *jh)
2530 {
2531         struct buffer_head *bh = jh2bh(jh);
2532
2533         jbd_lock_bh_journal_head(bh);
2534         J_ASSERT_JH(jh, jh->b_jcount > 0);
2535         --jh->b_jcount;
2536         if (!jh->b_jcount) {
2537                 __journal_remove_journal_head(bh);
2538                 jbd_unlock_bh_journal_head(bh);
2539                 __brelse(bh);
2540         } else
2541                 jbd_unlock_bh_journal_head(bh);
2542 }
2543
2544 /*
2545  * Initialize jbd inode head
2546  */
2547 void jbd2_journal_init_jbd_inode(struct jbd2_inode *jinode, struct inode *inode)
2548 {
2549         jinode->i_transaction = NULL;
2550         jinode->i_next_transaction = NULL;
2551         jinode->i_vfs_inode = inode;
2552         jinode->i_flags = 0;
2553         INIT_LIST_HEAD(&jinode->i_list);
2554 }
2555
2556 /*
2557  * Function to be called before we start removing inode from memory (i.e.,
2558  * clear_inode() is a fine place to be called from). It removes inode from
2559  * transaction's lists.
2560  */
2561 void jbd2_journal_release_jbd_inode(journal_t *journal,
2562                                     struct jbd2_inode *jinode)
2563 {
2564         if (!journal)
2565                 return;
2566 restart:
2567         spin_lock(&journal->j_list_lock);
2568         /* Is commit writing out inode - we have to wait */
2569         if (jinode->i_flags & JI_COMMIT_RUNNING) {
2570                 wait_queue_head_t *wq;
2571                 DEFINE_WAIT_BIT(wait, &jinode->i_flags, __JI_COMMIT_RUNNING);
2572                 wq = bit_waitqueue(&jinode->i_flags, __JI_COMMIT_RUNNING);
2573                 prepare_to_wait(wq, &wait.wait, TASK_UNINTERRUPTIBLE);
2574                 spin_unlock(&journal->j_list_lock);
2575                 schedule();
2576                 finish_wait(wq, &wait.wait);
2577                 goto restart;
2578         }
2579
2580         if (jinode->i_transaction) {
2581                 list_del(&jinode->i_list);
2582                 jinode->i_transaction = NULL;
2583         }
2584         spin_unlock(&journal->j_list_lock);
2585 }
2586
2587
2588 #ifdef CONFIG_PROC_FS
2589
2590 #define JBD2_STATS_PROC_NAME "fs/jbd2"
2591
2592 static void __init jbd2_create_jbd_stats_proc_entry(void)
2593 {
2594         proc_jbd2_stats = proc_mkdir(JBD2_STATS_PROC_NAME, NULL);
2595 }
2596
2597 static void __exit jbd2_remove_jbd_stats_proc_entry(void)
2598 {
2599         if (proc_jbd2_stats)
2600                 remove_proc_entry(JBD2_STATS_PROC_NAME, NULL);
2601 }
2602
2603 #else
2604
2605 #define jbd2_create_jbd_stats_proc_entry() do {} while (0)
2606 #define jbd2_remove_jbd_stats_proc_entry() do {} while (0)
2607
2608 #endif
2609
2610 struct kmem_cache *jbd2_handle_cache, *jbd2_inode_cache;
2611
2612 static int __init jbd2_journal_init_handle_cache(void)
2613 {
2614         jbd2_handle_cache = KMEM_CACHE(jbd2_journal_handle, SLAB_TEMPORARY);
2615         if (jbd2_handle_cache == NULL) {
2616                 printk(KERN_EMERG "JBD2: failed to create handle cache\n");
2617                 return -ENOMEM;
2618         }
2619         jbd2_inode_cache = KMEM_CACHE(jbd2_inode, 0);
2620         if (jbd2_inode_cache == NULL) {
2621                 printk(KERN_EMERG "JBD2: failed to create inode cache\n");
2622                 kmem_cache_destroy(jbd2_handle_cache);
2623                 return -ENOMEM;
2624         }
2625         return 0;
2626 }
2627
2628 static void jbd2_journal_destroy_handle_cache(void)
2629 {
2630         if (jbd2_handle_cache)
2631                 kmem_cache_destroy(jbd2_handle_cache);
2632         if (jbd2_inode_cache)
2633                 kmem_cache_destroy(jbd2_inode_cache);
2634
2635 }
2636
2637 /*
2638  * Module startup and shutdown
2639  */
2640
2641 static int __init journal_init_caches(void)
2642 {
2643         int ret;
2644
2645         ret = jbd2_journal_init_revoke_caches();
2646         if (ret == 0)
2647                 ret = jbd2_journal_init_journal_head_cache();
2648         if (ret == 0)
2649                 ret = jbd2_journal_init_handle_cache();
2650         if (ret == 0)
2651                 ret = jbd2_journal_init_transaction_cache();
2652         return ret;
2653 }
2654
2655 static void jbd2_journal_destroy_caches(void)
2656 {
2657         jbd2_journal_destroy_revoke_caches();
2658         jbd2_journal_destroy_journal_head_cache();
2659         jbd2_journal_destroy_handle_cache();
2660         jbd2_journal_destroy_transaction_cache();
2661         jbd2_journal_destroy_slabs();
2662 }
2663
2664 static int __init journal_init(void)
2665 {
2666         int ret;
2667
2668         BUILD_BUG_ON(sizeof(struct journal_superblock_s) != 1024);
2669
2670         ret = journal_init_caches();
2671         if (ret == 0) {
2672                 jbd2_create_jbd_stats_proc_entry();
2673         } else {
2674                 jbd2_journal_destroy_caches();
2675         }
2676         return ret;
2677 }
2678
2679 static void __exit journal_exit(void)
2680 {
2681 #ifdef CONFIG_JBD2_DEBUG
2682         int n = atomic_read(&nr_journal_heads);
2683         if (n)
2684                 printk(KERN_ERR "JBD2: leaked %d journal_heads!\n", n);
2685 #endif
2686         jbd2_remove_jbd_stats_proc_entry();
2687         jbd2_journal_destroy_caches();
2688 }
2689
2690 MODULE_LICENSE("GPL");
2691 module_init(journal_init);
2692 module_exit(journal_exit);
2693