Merge git://git.kernel.org/pub/scm/linux/kernel/git/rusty/linux-2.6-module-and-param
[linux-2.6] / fs / jbd / journal.c
1 /*
2  * linux/fs/jbd/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/jbd.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/debugfs.h>
39
40 #include <asm/uaccess.h>
41 #include <asm/page.h>
42
43 EXPORT_SYMBOL(journal_start);
44 EXPORT_SYMBOL(journal_restart);
45 EXPORT_SYMBOL(journal_extend);
46 EXPORT_SYMBOL(journal_stop);
47 EXPORT_SYMBOL(journal_lock_updates);
48 EXPORT_SYMBOL(journal_unlock_updates);
49 EXPORT_SYMBOL(journal_get_write_access);
50 EXPORT_SYMBOL(journal_get_create_access);
51 EXPORT_SYMBOL(journal_get_undo_access);
52 EXPORT_SYMBOL(journal_dirty_data);
53 EXPORT_SYMBOL(journal_dirty_metadata);
54 EXPORT_SYMBOL(journal_release_buffer);
55 EXPORT_SYMBOL(journal_forget);
56 #if 0
57 EXPORT_SYMBOL(journal_sync_buffer);
58 #endif
59 EXPORT_SYMBOL(journal_flush);
60 EXPORT_SYMBOL(journal_revoke);
61
62 EXPORT_SYMBOL(journal_init_dev);
63 EXPORT_SYMBOL(journal_init_inode);
64 EXPORT_SYMBOL(journal_update_format);
65 EXPORT_SYMBOL(journal_check_used_features);
66 EXPORT_SYMBOL(journal_check_available_features);
67 EXPORT_SYMBOL(journal_set_features);
68 EXPORT_SYMBOL(journal_create);
69 EXPORT_SYMBOL(journal_load);
70 EXPORT_SYMBOL(journal_destroy);
71 EXPORT_SYMBOL(journal_abort);
72 EXPORT_SYMBOL(journal_errno);
73 EXPORT_SYMBOL(journal_ack_err);
74 EXPORT_SYMBOL(journal_clear_err);
75 EXPORT_SYMBOL(log_wait_commit);
76 EXPORT_SYMBOL(journal_start_commit);
77 EXPORT_SYMBOL(journal_force_commit_nested);
78 EXPORT_SYMBOL(journal_wipe);
79 EXPORT_SYMBOL(journal_blocks_per_page);
80 EXPORT_SYMBOL(journal_invalidatepage);
81 EXPORT_SYMBOL(journal_try_to_free_buffers);
82 EXPORT_SYMBOL(journal_force_commit);
83
84 static int journal_convert_superblock_v1(journal_t *, journal_superblock_t *);
85 static void __journal_abort_soft (journal_t *journal, int errno);
86
87 /*
88  * Helper function used to manage commit timeouts
89  */
90
91 static void commit_timeout(unsigned long __data)
92 {
93         struct task_struct * p = (struct task_struct *) __data;
94
95         wake_up_process(p);
96 }
97
98 /*
99  * kjournald: The main thread function used to manage a logging device
100  * journal.
101  *
102  * This kernel thread is responsible for two things:
103  *
104  * 1) COMMIT:  Every so often we need to commit the current state of the
105  *    filesystem to disk.  The journal thread is responsible for writing
106  *    all of the metadata buffers to disk.
107  *
108  * 2) CHECKPOINT: We cannot reuse a used section of the log file until all
109  *    of the data in that part of the log has been rewritten elsewhere on
110  *    the disk.  Flushing these old buffers to reclaim space in the log is
111  *    known as checkpointing, and this thread is responsible for that job.
112  */
113
114 static int kjournald(void *arg)
115 {
116         journal_t *journal = arg;
117         transaction_t *transaction;
118
119         /*
120          * Set up an interval timer which can be used to trigger a commit wakeup
121          * after the commit interval expires
122          */
123         setup_timer(&journal->j_commit_timer, commit_timeout,
124                         (unsigned long)current);
125
126         /* Record that the journal thread is running */
127         journal->j_task = current;
128         wake_up(&journal->j_wait_done_commit);
129
130         printk(KERN_INFO "kjournald starting.  Commit interval %ld seconds\n",
131                         journal->j_commit_interval / HZ);
132
133         /*
134          * And now, wait forever for commit wakeup events.
135          */
136         spin_lock(&journal->j_state_lock);
137
138 loop:
139         if (journal->j_flags & JFS_UNMOUNT)
140                 goto end_loop;
141
142         jbd_debug(1, "commit_sequence=%d, commit_request=%d\n",
143                 journal->j_commit_sequence, journal->j_commit_request);
144
145         if (journal->j_commit_sequence != journal->j_commit_request) {
146                 jbd_debug(1, "OK, requests differ\n");
147                 spin_unlock(&journal->j_state_lock);
148                 del_timer_sync(&journal->j_commit_timer);
149                 journal_commit_transaction(journal);
150                 spin_lock(&journal->j_state_lock);
151                 goto loop;
152         }
153
154         wake_up(&journal->j_wait_done_commit);
155         if (freezing(current)) {
156                 /*
157                  * The simpler the better. Flushing journal isn't a
158                  * good idea, because that depends on threads that may
159                  * be already stopped.
160                  */
161                 jbd_debug(1, "Now suspending kjournald\n");
162                 spin_unlock(&journal->j_state_lock);
163                 refrigerator();
164                 spin_lock(&journal->j_state_lock);
165         } else {
166                 /*
167                  * We assume on resume that commits are already there,
168                  * so we don't sleep
169                  */
170                 DEFINE_WAIT(wait);
171                 int should_sleep = 1;
172
173                 prepare_to_wait(&journal->j_wait_commit, &wait,
174                                 TASK_INTERRUPTIBLE);
175                 if (journal->j_commit_sequence != journal->j_commit_request)
176                         should_sleep = 0;
177                 transaction = journal->j_running_transaction;
178                 if (transaction && time_after_eq(jiffies,
179                                                 transaction->t_expires))
180                         should_sleep = 0;
181                 if (journal->j_flags & JFS_UNMOUNT)
182                         should_sleep = 0;
183                 if (should_sleep) {
184                         spin_unlock(&journal->j_state_lock);
185                         schedule();
186                         spin_lock(&journal->j_state_lock);
187                 }
188                 finish_wait(&journal->j_wait_commit, &wait);
189         }
190
191         jbd_debug(1, "kjournald wakes\n");
192
193         /*
194          * Were we woken up by a commit wakeup event?
195          */
196         transaction = journal->j_running_transaction;
197         if (transaction && time_after_eq(jiffies, transaction->t_expires)) {
198                 journal->j_commit_request = transaction->t_tid;
199                 jbd_debug(1, "woke because of timeout\n");
200         }
201         goto loop;
202
203 end_loop:
204         spin_unlock(&journal->j_state_lock);
205         del_timer_sync(&journal->j_commit_timer);
206         journal->j_task = NULL;
207         wake_up(&journal->j_wait_done_commit);
208         jbd_debug(1, "Journal thread exiting.\n");
209         return 0;
210 }
211
212 static int journal_start_thread(journal_t *journal)
213 {
214         struct task_struct *t;
215
216         t = kthread_run(kjournald, journal, "kjournald");
217         if (IS_ERR(t))
218                 return PTR_ERR(t);
219
220         wait_event(journal->j_wait_done_commit, journal->j_task != NULL);
221         return 0;
222 }
223
224 static void journal_kill_thread(journal_t *journal)
225 {
226         spin_lock(&journal->j_state_lock);
227         journal->j_flags |= JFS_UNMOUNT;
228
229         while (journal->j_task) {
230                 wake_up(&journal->j_wait_commit);
231                 spin_unlock(&journal->j_state_lock);
232                 wait_event(journal->j_wait_done_commit,
233                                 journal->j_task == NULL);
234                 spin_lock(&journal->j_state_lock);
235         }
236         spin_unlock(&journal->j_state_lock);
237 }
238
239 /*
240  * journal_write_metadata_buffer: write a metadata buffer to the journal.
241  *
242  * Writes a metadata buffer to a given disk block.  The actual IO is not
243  * performed but a new buffer_head is constructed which labels the data
244  * to be written with the correct destination disk block.
245  *
246  * Any magic-number escaping which needs to be done will cause a
247  * copy-out here.  If the buffer happens to start with the
248  * JFS_MAGIC_NUMBER, then we can't write it to the log directly: the
249  * magic number is only written to the log for descripter blocks.  In
250  * this case, we copy the data and replace the first word with 0, and we
251  * return a result code which indicates that this buffer needs to be
252  * marked as an escaped buffer in the corresponding log descriptor
253  * block.  The missing word can then be restored when the block is read
254  * during recovery.
255  *
256  * If the source buffer has already been modified by a new transaction
257  * since we took the last commit snapshot, we use the frozen copy of
258  * that data for IO.  If we end up using the existing buffer_head's data
259  * for the write, then we *have* to lock the buffer to prevent anyone
260  * else from using and possibly modifying it while the IO is in
261  * progress.
262  *
263  * The function returns a pointer to the buffer_heads to be used for IO.
264  *
265  * We assume that the journal has already been locked in this function.
266  *
267  * Return value:
268  *  <0: Error
269  * >=0: Finished OK
270  *
271  * On success:
272  * Bit 0 set == escape performed on the data
273  * Bit 1 set == buffer copy-out performed (kfree the data after IO)
274  */
275
276 int journal_write_metadata_buffer(transaction_t *transaction,
277                                   struct journal_head  *jh_in,
278                                   struct journal_head **jh_out,
279                                   unsigned long blocknr)
280 {
281         int need_copy_out = 0;
282         int done_copy_out = 0;
283         int do_escape = 0;
284         char *mapped_data;
285         struct buffer_head *new_bh;
286         struct journal_head *new_jh;
287         struct page *new_page;
288         unsigned int new_offset;
289         struct buffer_head *bh_in = jh2bh(jh_in);
290
291         /*
292          * The buffer really shouldn't be locked: only the current committing
293          * transaction is allowed to write it, so nobody else is allowed
294          * to do any IO.
295          *
296          * akpm: except if we're journalling data, and write() output is
297          * also part of a shared mapping, and another thread has
298          * decided to launch a writepage() against this buffer.
299          */
300         J_ASSERT_BH(bh_in, buffer_jbddirty(bh_in));
301
302         new_bh = alloc_buffer_head(GFP_NOFS|__GFP_NOFAIL);
303
304         /*
305          * If a new transaction has already done a buffer copy-out, then
306          * we use that version of the data for the commit.
307          */
308         jbd_lock_bh_state(bh_in);
309 repeat:
310         if (jh_in->b_frozen_data) {
311                 done_copy_out = 1;
312                 new_page = virt_to_page(jh_in->b_frozen_data);
313                 new_offset = offset_in_page(jh_in->b_frozen_data);
314         } else {
315                 new_page = jh2bh(jh_in)->b_page;
316                 new_offset = offset_in_page(jh2bh(jh_in)->b_data);
317         }
318
319         mapped_data = kmap_atomic(new_page, KM_USER0);
320         /*
321          * Check for escaping
322          */
323         if (*((__be32 *)(mapped_data + new_offset)) ==
324                                 cpu_to_be32(JFS_MAGIC_NUMBER)) {
325                 need_copy_out = 1;
326                 do_escape = 1;
327         }
328         kunmap_atomic(mapped_data, KM_USER0);
329
330         /*
331          * Do we need to do a data copy?
332          */
333         if (need_copy_out && !done_copy_out) {
334                 char *tmp;
335
336                 jbd_unlock_bh_state(bh_in);
337                 tmp = jbd_alloc(bh_in->b_size, GFP_NOFS);
338                 jbd_lock_bh_state(bh_in);
339                 if (jh_in->b_frozen_data) {
340                         jbd_free(tmp, bh_in->b_size);
341                         goto repeat;
342                 }
343
344                 jh_in->b_frozen_data = tmp;
345                 mapped_data = kmap_atomic(new_page, KM_USER0);
346                 memcpy(tmp, mapped_data + new_offset, jh2bh(jh_in)->b_size);
347                 kunmap_atomic(mapped_data, KM_USER0);
348
349                 new_page = virt_to_page(tmp);
350                 new_offset = offset_in_page(tmp);
351                 done_copy_out = 1;
352         }
353
354         /*
355          * Did we need to do an escaping?  Now we've done all the
356          * copying, we can finally do so.
357          */
358         if (do_escape) {
359                 mapped_data = kmap_atomic(new_page, KM_USER0);
360                 *((unsigned int *)(mapped_data + new_offset)) = 0;
361                 kunmap_atomic(mapped_data, KM_USER0);
362         }
363
364         /* keep subsequent assertions sane */
365         new_bh->b_state = 0;
366         init_buffer(new_bh, NULL, NULL);
367         atomic_set(&new_bh->b_count, 1);
368         jbd_unlock_bh_state(bh_in);
369
370         new_jh = journal_add_journal_head(new_bh);      /* This sleeps */
371
372         set_bh_page(new_bh, new_page, new_offset);
373         new_jh->b_transaction = NULL;
374         new_bh->b_size = jh2bh(jh_in)->b_size;
375         new_bh->b_bdev = transaction->t_journal->j_dev;
376         new_bh->b_blocknr = blocknr;
377         set_buffer_mapped(new_bh);
378         set_buffer_dirty(new_bh);
379
380         *jh_out = new_jh;
381
382         /*
383          * The to-be-written buffer needs to get moved to the io queue,
384          * and the original buffer whose contents we are shadowing or
385          * copying is moved to the transaction's shadow queue.
386          */
387         JBUFFER_TRACE(jh_in, "file as BJ_Shadow");
388         journal_file_buffer(jh_in, transaction, BJ_Shadow);
389         JBUFFER_TRACE(new_jh, "file as BJ_IO");
390         journal_file_buffer(new_jh, transaction, BJ_IO);
391
392         return do_escape | (done_copy_out << 1);
393 }
394
395 /*
396  * Allocation code for the journal file.  Manage the space left in the
397  * journal, so that we can begin checkpointing when appropriate.
398  */
399
400 /*
401  * __log_space_left: Return the number of free blocks left in the journal.
402  *
403  * Called with the journal already locked.
404  *
405  * Called under j_state_lock
406  */
407
408 int __log_space_left(journal_t *journal)
409 {
410         int left = journal->j_free;
411
412         assert_spin_locked(&journal->j_state_lock);
413
414         /*
415          * Be pessimistic here about the number of those free blocks which
416          * might be required for log descriptor control blocks.
417          */
418
419 #define MIN_LOG_RESERVED_BLOCKS 32 /* Allow for rounding errors */
420
421         left -= MIN_LOG_RESERVED_BLOCKS;
422
423         if (left <= 0)
424                 return 0;
425         left -= (left >> 3);
426         return left;
427 }
428
429 /*
430  * Called under j_state_lock.  Returns true if a transaction commit was started.
431  */
432 int __log_start_commit(journal_t *journal, tid_t target)
433 {
434         /*
435          * Are we already doing a recent enough commit?
436          */
437         if (!tid_geq(journal->j_commit_request, target)) {
438                 /*
439                  * We want a new commit: OK, mark the request and wakup the
440                  * commit thread.  We do _not_ do the commit ourselves.
441                  */
442
443                 journal->j_commit_request = target;
444                 jbd_debug(1, "JBD: requesting commit %d/%d\n",
445                           journal->j_commit_request,
446                           journal->j_commit_sequence);
447                 wake_up(&journal->j_wait_commit);
448                 return 1;
449         }
450         return 0;
451 }
452
453 int log_start_commit(journal_t *journal, tid_t tid)
454 {
455         int ret;
456
457         spin_lock(&journal->j_state_lock);
458         ret = __log_start_commit(journal, tid);
459         spin_unlock(&journal->j_state_lock);
460         return ret;
461 }
462
463 /*
464  * Force and wait upon a commit if the calling process is not within
465  * transaction.  This is used for forcing out undo-protected data which contains
466  * bitmaps, when the fs is running out of space.
467  *
468  * We can only force the running transaction if we don't have an active handle;
469  * otherwise, we will deadlock.
470  *
471  * Returns true if a transaction was started.
472  */
473 int journal_force_commit_nested(journal_t *journal)
474 {
475         transaction_t *transaction = NULL;
476         tid_t tid;
477
478         spin_lock(&journal->j_state_lock);
479         if (journal->j_running_transaction && !current->journal_info) {
480                 transaction = journal->j_running_transaction;
481                 __log_start_commit(journal, transaction->t_tid);
482         } else if (journal->j_committing_transaction)
483                 transaction = journal->j_committing_transaction;
484
485         if (!transaction) {
486                 spin_unlock(&journal->j_state_lock);
487                 return 0;       /* Nothing to retry */
488         }
489
490         tid = transaction->t_tid;
491         spin_unlock(&journal->j_state_lock);
492         log_wait_commit(journal, tid);
493         return 1;
494 }
495
496 /*
497  * Start a commit of the current running transaction (if any).  Returns true
498  * if a transaction is going to be committed (or is currently already
499  * committing), and fills its tid in at *ptid
500  */
501 int journal_start_commit(journal_t *journal, tid_t *ptid)
502 {
503         int ret = 0;
504
505         spin_lock(&journal->j_state_lock);
506         if (journal->j_running_transaction) {
507                 tid_t tid = journal->j_running_transaction->t_tid;
508
509                 __log_start_commit(journal, tid);
510                 /* There's a running transaction and we've just made sure
511                  * it's commit has been scheduled. */
512                 if (ptid)
513                         *ptid = tid;
514                 ret = 1;
515         } else if (journal->j_committing_transaction) {
516                 /*
517                  * If ext3_write_super() recently started a commit, then we
518                  * have to wait for completion of that transaction
519                  */
520                 if (ptid)
521                         *ptid = journal->j_committing_transaction->t_tid;
522                 ret = 1;
523         }
524         spin_unlock(&journal->j_state_lock);
525         return ret;
526 }
527
528 /*
529  * Wait for a specified commit to complete.
530  * The caller may not hold the journal lock.
531  */
532 int log_wait_commit(journal_t *journal, tid_t tid)
533 {
534         int err = 0;
535
536 #ifdef CONFIG_JBD_DEBUG
537         spin_lock(&journal->j_state_lock);
538         if (!tid_geq(journal->j_commit_request, tid)) {
539                 printk(KERN_EMERG
540                        "%s: error: j_commit_request=%d, tid=%d\n",
541                        __func__, journal->j_commit_request, tid);
542         }
543         spin_unlock(&journal->j_state_lock);
544 #endif
545         spin_lock(&journal->j_state_lock);
546         while (tid_gt(tid, journal->j_commit_sequence)) {
547                 jbd_debug(1, "JBD: want %d, j_commit_sequence=%d\n",
548                                   tid, journal->j_commit_sequence);
549                 wake_up(&journal->j_wait_commit);
550                 spin_unlock(&journal->j_state_lock);
551                 wait_event(journal->j_wait_done_commit,
552                                 !tid_gt(tid, journal->j_commit_sequence));
553                 spin_lock(&journal->j_state_lock);
554         }
555         spin_unlock(&journal->j_state_lock);
556
557         if (unlikely(is_journal_aborted(journal))) {
558                 printk(KERN_EMERG "journal commit I/O error\n");
559                 err = -EIO;
560         }
561         return err;
562 }
563
564 /*
565  * Log buffer allocation routines:
566  */
567
568 int journal_next_log_block(journal_t *journal, unsigned long *retp)
569 {
570         unsigned long blocknr;
571
572         spin_lock(&journal->j_state_lock);
573         J_ASSERT(journal->j_free > 1);
574
575         blocknr = journal->j_head;
576         journal->j_head++;
577         journal->j_free--;
578         if (journal->j_head == journal->j_last)
579                 journal->j_head = journal->j_first;
580         spin_unlock(&journal->j_state_lock);
581         return journal_bmap(journal, blocknr, retp);
582 }
583
584 /*
585  * Conversion of logical to physical block numbers for the journal
586  *
587  * On external journals the journal blocks are identity-mapped, so
588  * this is a no-op.  If needed, we can use j_blk_offset - everything is
589  * ready.
590  */
591 int journal_bmap(journal_t *journal, unsigned long blocknr,
592                  unsigned long *retp)
593 {
594         int err = 0;
595         unsigned long ret;
596
597         if (journal->j_inode) {
598                 ret = bmap(journal->j_inode, blocknr);
599                 if (ret)
600                         *retp = ret;
601                 else {
602                         char b[BDEVNAME_SIZE];
603
604                         printk(KERN_ALERT "%s: journal block not found "
605                                         "at offset %lu on %s\n",
606                                 __func__,
607                                 blocknr,
608                                 bdevname(journal->j_dev, b));
609                         err = -EIO;
610                         __journal_abort_soft(journal, err);
611                 }
612         } else {
613                 *retp = blocknr; /* +journal->j_blk_offset */
614         }
615         return err;
616 }
617
618 /*
619  * We play buffer_head aliasing tricks to write data/metadata blocks to
620  * the journal without copying their contents, but for journal
621  * descriptor blocks we do need to generate bona fide buffers.
622  *
623  * After the caller of journal_get_descriptor_buffer() has finished modifying
624  * the buffer's contents they really should run flush_dcache_page(bh->b_page).
625  * But we don't bother doing that, so there will be coherency problems with
626  * mmaps of blockdevs which hold live JBD-controlled filesystems.
627  */
628 struct journal_head *journal_get_descriptor_buffer(journal_t *journal)
629 {
630         struct buffer_head *bh;
631         unsigned long blocknr;
632         int err;
633
634         err = journal_next_log_block(journal, &blocknr);
635
636         if (err)
637                 return NULL;
638
639         bh = __getblk(journal->j_dev, blocknr, journal->j_blocksize);
640         if (!bh)
641                 return NULL;
642         lock_buffer(bh);
643         memset(bh->b_data, 0, journal->j_blocksize);
644         set_buffer_uptodate(bh);
645         unlock_buffer(bh);
646         BUFFER_TRACE(bh, "return this buffer");
647         return journal_add_journal_head(bh);
648 }
649
650 /*
651  * Management for journal control blocks: functions to create and
652  * destroy journal_t structures, and to initialise and read existing
653  * journal blocks from disk.  */
654
655 /* First: create and setup a journal_t object in memory.  We initialise
656  * very few fields yet: that has to wait until we have created the
657  * journal structures from from scratch, or loaded them from disk. */
658
659 static journal_t * journal_init_common (void)
660 {
661         journal_t *journal;
662         int err;
663
664         journal = kzalloc(sizeof(*journal), GFP_KERNEL);
665         if (!journal)
666                 goto fail;
667
668         init_waitqueue_head(&journal->j_wait_transaction_locked);
669         init_waitqueue_head(&journal->j_wait_logspace);
670         init_waitqueue_head(&journal->j_wait_done_commit);
671         init_waitqueue_head(&journal->j_wait_checkpoint);
672         init_waitqueue_head(&journal->j_wait_commit);
673         init_waitqueue_head(&journal->j_wait_updates);
674         mutex_init(&journal->j_barrier);
675         mutex_init(&journal->j_checkpoint_mutex);
676         spin_lock_init(&journal->j_revoke_lock);
677         spin_lock_init(&journal->j_list_lock);
678         spin_lock_init(&journal->j_state_lock);
679
680         journal->j_commit_interval = (HZ * JBD_DEFAULT_MAX_COMMIT_AGE);
681
682         /* The journal is marked for error until we succeed with recovery! */
683         journal->j_flags = JFS_ABORT;
684
685         /* Set up a default-sized revoke table for the new mount. */
686         err = journal_init_revoke(journal, JOURNAL_REVOKE_DEFAULT_HASH);
687         if (err) {
688                 kfree(journal);
689                 goto fail;
690         }
691         return journal;
692 fail:
693         return NULL;
694 }
695
696 /* journal_init_dev and journal_init_inode:
697  *
698  * Create a journal structure assigned some fixed set of disk blocks to
699  * the journal.  We don't actually touch those disk blocks yet, but we
700  * need to set up all of the mapping information to tell the journaling
701  * system where the journal blocks are.
702  *
703  */
704
705 /**
706  *  journal_t * journal_init_dev() - creates and initialises a journal structure
707  *  @bdev: Block device on which to create the journal
708  *  @fs_dev: Device which hold journalled filesystem for this journal.
709  *  @start: Block nr Start of journal.
710  *  @len:  Length of the journal in blocks.
711  *  @blocksize: blocksize of journalling device
712  *
713  *  Returns: a newly created journal_t *
714  *
715  *  journal_init_dev creates a journal which maps a fixed contiguous
716  *  range of blocks on an arbitrary block device.
717  *
718  */
719 journal_t * journal_init_dev(struct block_device *bdev,
720                         struct block_device *fs_dev,
721                         int start, int len, int blocksize)
722 {
723         journal_t *journal = journal_init_common();
724         struct buffer_head *bh;
725         int n;
726
727         if (!journal)
728                 return NULL;
729
730         /* journal descriptor can store up to n blocks -bzzz */
731         journal->j_blocksize = blocksize;
732         n = journal->j_blocksize / sizeof(journal_block_tag_t);
733         journal->j_wbufsize = n;
734         journal->j_wbuf = kmalloc(n * sizeof(struct buffer_head*), GFP_KERNEL);
735         if (!journal->j_wbuf) {
736                 printk(KERN_ERR "%s: Cant allocate bhs for commit thread\n",
737                         __func__);
738                 goto out_err;
739         }
740         journal->j_dev = bdev;
741         journal->j_fs_dev = fs_dev;
742         journal->j_blk_offset = start;
743         journal->j_maxlen = len;
744
745         bh = __getblk(journal->j_dev, start, journal->j_blocksize);
746         if (!bh) {
747                 printk(KERN_ERR
748                        "%s: Cannot get buffer for journal superblock\n",
749                        __func__);
750                 goto out_err;
751         }
752         journal->j_sb_buffer = bh;
753         journal->j_superblock = (journal_superblock_t *)bh->b_data;
754
755         return journal;
756 out_err:
757         kfree(journal);
758         return NULL;
759 }
760
761 /**
762  *  journal_t * journal_init_inode () - creates a journal which maps to a inode.
763  *  @inode: An inode to create the journal in
764  *
765  * journal_init_inode creates a journal which maps an on-disk inode as
766  * the journal.  The inode must exist already, must support bmap() and
767  * must have all data blocks preallocated.
768  */
769 journal_t * journal_init_inode (struct inode *inode)
770 {
771         struct buffer_head *bh;
772         journal_t *journal = journal_init_common();
773         int err;
774         int n;
775         unsigned long blocknr;
776
777         if (!journal)
778                 return NULL;
779
780         journal->j_dev = journal->j_fs_dev = inode->i_sb->s_bdev;
781         journal->j_inode = inode;
782         jbd_debug(1,
783                   "journal %p: inode %s/%ld, size %Ld, bits %d, blksize %ld\n",
784                   journal, inode->i_sb->s_id, inode->i_ino,
785                   (long long) inode->i_size,
786                   inode->i_sb->s_blocksize_bits, inode->i_sb->s_blocksize);
787
788         journal->j_maxlen = inode->i_size >> inode->i_sb->s_blocksize_bits;
789         journal->j_blocksize = inode->i_sb->s_blocksize;
790
791         /* journal descriptor can store up to n blocks -bzzz */
792         n = journal->j_blocksize / sizeof(journal_block_tag_t);
793         journal->j_wbufsize = n;
794         journal->j_wbuf = kmalloc(n * sizeof(struct buffer_head*), GFP_KERNEL);
795         if (!journal->j_wbuf) {
796                 printk(KERN_ERR "%s: Cant allocate bhs for commit thread\n",
797                         __func__);
798                 goto out_err;
799         }
800
801         err = journal_bmap(journal, 0, &blocknr);
802         /* If that failed, give up */
803         if (err) {
804                 printk(KERN_ERR "%s: Cannnot locate journal superblock\n",
805                        __func__);
806                 goto out_err;
807         }
808
809         bh = __getblk(journal->j_dev, blocknr, journal->j_blocksize);
810         if (!bh) {
811                 printk(KERN_ERR
812                        "%s: Cannot get buffer for journal superblock\n",
813                        __func__);
814                 goto out_err;
815         }
816         journal->j_sb_buffer = bh;
817         journal->j_superblock = (journal_superblock_t *)bh->b_data;
818
819         return journal;
820 out_err:
821         kfree(journal);
822         return NULL;
823 }
824
825 /*
826  * If the journal init or create aborts, we need to mark the journal
827  * superblock as being NULL to prevent the journal destroy from writing
828  * back a bogus superblock.
829  */
830 static void journal_fail_superblock (journal_t *journal)
831 {
832         struct buffer_head *bh = journal->j_sb_buffer;
833         brelse(bh);
834         journal->j_sb_buffer = NULL;
835 }
836
837 /*
838  * Given a journal_t structure, initialise the various fields for
839  * startup of a new journaling session.  We use this both when creating
840  * a journal, and after recovering an old journal to reset it for
841  * subsequent use.
842  */
843
844 static int journal_reset(journal_t *journal)
845 {
846         journal_superblock_t *sb = journal->j_superblock;
847         unsigned long first, last;
848
849         first = be32_to_cpu(sb->s_first);
850         last = be32_to_cpu(sb->s_maxlen);
851
852         journal->j_first = first;
853         journal->j_last = last;
854
855         journal->j_head = first;
856         journal->j_tail = first;
857         journal->j_free = last - first;
858
859         journal->j_tail_sequence = journal->j_transaction_sequence;
860         journal->j_commit_sequence = journal->j_transaction_sequence - 1;
861         journal->j_commit_request = journal->j_commit_sequence;
862
863         journal->j_max_transaction_buffers = journal->j_maxlen / 4;
864
865         /* Add the dynamic fields and write it to disk. */
866         journal_update_superblock(journal, 1);
867         return journal_start_thread(journal);
868 }
869
870 /**
871  * int journal_create() - Initialise the new journal file
872  * @journal: Journal to create. This structure must have been initialised
873  *
874  * Given a journal_t structure which tells us which disk blocks we can
875  * use, create a new journal superblock and initialise all of the
876  * journal fields from scratch.
877  **/
878 int journal_create(journal_t *journal)
879 {
880         unsigned long blocknr;
881         struct buffer_head *bh;
882         journal_superblock_t *sb;
883         int i, err;
884
885         if (journal->j_maxlen < JFS_MIN_JOURNAL_BLOCKS) {
886                 printk (KERN_ERR "Journal length (%d blocks) too short.\n",
887                         journal->j_maxlen);
888                 journal_fail_superblock(journal);
889                 return -EINVAL;
890         }
891
892         if (journal->j_inode == NULL) {
893                 /*
894                  * We don't know what block to start at!
895                  */
896                 printk(KERN_EMERG
897                        "%s: creation of journal on external device!\n",
898                        __func__);
899                 BUG();
900         }
901
902         /* Zero out the entire journal on disk.  We cannot afford to
903            have any blocks on disk beginning with JFS_MAGIC_NUMBER. */
904         jbd_debug(1, "JBD: Zeroing out journal blocks...\n");
905         for (i = 0; i < journal->j_maxlen; i++) {
906                 err = journal_bmap(journal, i, &blocknr);
907                 if (err)
908                         return err;
909                 bh = __getblk(journal->j_dev, blocknr, journal->j_blocksize);
910                 lock_buffer(bh);
911                 memset (bh->b_data, 0, journal->j_blocksize);
912                 BUFFER_TRACE(bh, "marking dirty");
913                 mark_buffer_dirty(bh);
914                 BUFFER_TRACE(bh, "marking uptodate");
915                 set_buffer_uptodate(bh);
916                 unlock_buffer(bh);
917                 __brelse(bh);
918         }
919
920         sync_blockdev(journal->j_dev);
921         jbd_debug(1, "JBD: journal cleared.\n");
922
923         /* OK, fill in the initial static fields in the new superblock */
924         sb = journal->j_superblock;
925
926         sb->s_header.h_magic     = cpu_to_be32(JFS_MAGIC_NUMBER);
927         sb->s_header.h_blocktype = cpu_to_be32(JFS_SUPERBLOCK_V2);
928
929         sb->s_blocksize = cpu_to_be32(journal->j_blocksize);
930         sb->s_maxlen    = cpu_to_be32(journal->j_maxlen);
931         sb->s_first     = cpu_to_be32(1);
932
933         journal->j_transaction_sequence = 1;
934
935         journal->j_flags &= ~JFS_ABORT;
936         journal->j_format_version = 2;
937
938         return journal_reset(journal);
939 }
940
941 /**
942  * void journal_update_superblock() - Update journal sb on disk.
943  * @journal: The journal to update.
944  * @wait: Set to '0' if you don't want to wait for IO completion.
945  *
946  * Update a journal's dynamic superblock fields and write it to disk,
947  * optionally waiting for the IO to complete.
948  */
949 void journal_update_superblock(journal_t *journal, int wait)
950 {
951         journal_superblock_t *sb = journal->j_superblock;
952         struct buffer_head *bh = journal->j_sb_buffer;
953
954         /*
955          * As a special case, if the on-disk copy is already marked as needing
956          * no recovery (s_start == 0) and there are no outstanding transactions
957          * in the filesystem, then we can safely defer the superblock update
958          * until the next commit by setting JFS_FLUSHED.  This avoids
959          * attempting a write to a potential-readonly device.
960          */
961         if (sb->s_start == 0 && journal->j_tail_sequence ==
962                                 journal->j_transaction_sequence) {
963                 jbd_debug(1,"JBD: Skipping superblock update on recovered sb "
964                         "(start %ld, seq %d, errno %d)\n",
965                         journal->j_tail, journal->j_tail_sequence,
966                         journal->j_errno);
967                 goto out;
968         }
969
970         spin_lock(&journal->j_state_lock);
971         jbd_debug(1,"JBD: updating superblock (start %ld, seq %d, errno %d)\n",
972                   journal->j_tail, journal->j_tail_sequence, journal->j_errno);
973
974         sb->s_sequence = cpu_to_be32(journal->j_tail_sequence);
975         sb->s_start    = cpu_to_be32(journal->j_tail);
976         sb->s_errno    = cpu_to_be32(journal->j_errno);
977         spin_unlock(&journal->j_state_lock);
978
979         BUFFER_TRACE(bh, "marking dirty");
980         mark_buffer_dirty(bh);
981         if (wait)
982                 sync_dirty_buffer(bh);
983         else
984                 ll_rw_block(SWRITE, 1, &bh);
985
986 out:
987         /* If we have just flushed the log (by marking s_start==0), then
988          * any future commit will have to be careful to update the
989          * superblock again to re-record the true start of the log. */
990
991         spin_lock(&journal->j_state_lock);
992         if (sb->s_start)
993                 journal->j_flags &= ~JFS_FLUSHED;
994         else
995                 journal->j_flags |= JFS_FLUSHED;
996         spin_unlock(&journal->j_state_lock);
997 }
998
999 /*
1000  * Read the superblock for a given journal, performing initial
1001  * validation of the format.
1002  */
1003
1004 static int journal_get_superblock(journal_t *journal)
1005 {
1006         struct buffer_head *bh;
1007         journal_superblock_t *sb;
1008         int err = -EIO;
1009
1010         bh = journal->j_sb_buffer;
1011
1012         J_ASSERT(bh != NULL);
1013         if (!buffer_uptodate(bh)) {
1014                 ll_rw_block(READ, 1, &bh);
1015                 wait_on_buffer(bh);
1016                 if (!buffer_uptodate(bh)) {
1017                         printk (KERN_ERR
1018                                 "JBD: IO error reading journal superblock\n");
1019                         goto out;
1020                 }
1021         }
1022
1023         sb = journal->j_superblock;
1024
1025         err = -EINVAL;
1026
1027         if (sb->s_header.h_magic != cpu_to_be32(JFS_MAGIC_NUMBER) ||
1028             sb->s_blocksize != cpu_to_be32(journal->j_blocksize)) {
1029                 printk(KERN_WARNING "JBD: no valid journal superblock found\n");
1030                 goto out;
1031         }
1032
1033         switch(be32_to_cpu(sb->s_header.h_blocktype)) {
1034         case JFS_SUPERBLOCK_V1:
1035                 journal->j_format_version = 1;
1036                 break;
1037         case JFS_SUPERBLOCK_V2:
1038                 journal->j_format_version = 2;
1039                 break;
1040         default:
1041                 printk(KERN_WARNING "JBD: unrecognised superblock format ID\n");
1042                 goto out;
1043         }
1044
1045         if (be32_to_cpu(sb->s_maxlen) < journal->j_maxlen)
1046                 journal->j_maxlen = be32_to_cpu(sb->s_maxlen);
1047         else if (be32_to_cpu(sb->s_maxlen) > journal->j_maxlen) {
1048                 printk (KERN_WARNING "JBD: journal file too short\n");
1049                 goto out;
1050         }
1051
1052         return 0;
1053
1054 out:
1055         journal_fail_superblock(journal);
1056         return err;
1057 }
1058
1059 /*
1060  * Load the on-disk journal superblock and read the key fields into the
1061  * journal_t.
1062  */
1063
1064 static int load_superblock(journal_t *journal)
1065 {
1066         int err;
1067         journal_superblock_t *sb;
1068
1069         err = journal_get_superblock(journal);
1070         if (err)
1071                 return err;
1072
1073         sb = journal->j_superblock;
1074
1075         journal->j_tail_sequence = be32_to_cpu(sb->s_sequence);
1076         journal->j_tail = be32_to_cpu(sb->s_start);
1077         journal->j_first = be32_to_cpu(sb->s_first);
1078         journal->j_last = be32_to_cpu(sb->s_maxlen);
1079         journal->j_errno = be32_to_cpu(sb->s_errno);
1080
1081         return 0;
1082 }
1083
1084
1085 /**
1086  * int journal_load() - Read journal from disk.
1087  * @journal: Journal to act on.
1088  *
1089  * Given a journal_t structure which tells us which disk blocks contain
1090  * a journal, read the journal from disk to initialise the in-memory
1091  * structures.
1092  */
1093 int journal_load(journal_t *journal)
1094 {
1095         int err;
1096         journal_superblock_t *sb;
1097
1098         err = load_superblock(journal);
1099         if (err)
1100                 return err;
1101
1102         sb = journal->j_superblock;
1103         /* If this is a V2 superblock, then we have to check the
1104          * features flags on it. */
1105
1106         if (journal->j_format_version >= 2) {
1107                 if ((sb->s_feature_ro_compat &
1108                      ~cpu_to_be32(JFS_KNOWN_ROCOMPAT_FEATURES)) ||
1109                     (sb->s_feature_incompat &
1110                      ~cpu_to_be32(JFS_KNOWN_INCOMPAT_FEATURES))) {
1111                         printk (KERN_WARNING
1112                                 "JBD: Unrecognised features on journal\n");
1113                         return -EINVAL;
1114                 }
1115         }
1116
1117         /* Let the recovery code check whether it needs to recover any
1118          * data from the journal. */
1119         if (journal_recover(journal))
1120                 goto recovery_error;
1121
1122         /* OK, we've finished with the dynamic journal bits:
1123          * reinitialise the dynamic contents of the superblock in memory
1124          * and reset them on disk. */
1125         if (journal_reset(journal))
1126                 goto recovery_error;
1127
1128         journal->j_flags &= ~JFS_ABORT;
1129         journal->j_flags |= JFS_LOADED;
1130         return 0;
1131
1132 recovery_error:
1133         printk (KERN_WARNING "JBD: recovery failed\n");
1134         return -EIO;
1135 }
1136
1137 /**
1138  * void journal_destroy() - Release a journal_t structure.
1139  * @journal: Journal to act on.
1140  *
1141  * Release a journal_t structure once it is no longer in use by the
1142  * journaled object.
1143  * Return <0 if we couldn't clean up the journal.
1144  */
1145 int journal_destroy(journal_t *journal)
1146 {
1147         int err = 0;
1148
1149         /* Wait for the commit thread to wake up and die. */
1150         journal_kill_thread(journal);
1151
1152         /* Force a final log commit */
1153         if (journal->j_running_transaction)
1154                 journal_commit_transaction(journal);
1155
1156         /* Force any old transactions to disk */
1157
1158         /* Totally anal locking here... */
1159         spin_lock(&journal->j_list_lock);
1160         while (journal->j_checkpoint_transactions != NULL) {
1161                 spin_unlock(&journal->j_list_lock);
1162                 log_do_checkpoint(journal);
1163                 spin_lock(&journal->j_list_lock);
1164         }
1165
1166         J_ASSERT(journal->j_running_transaction == NULL);
1167         J_ASSERT(journal->j_committing_transaction == NULL);
1168         J_ASSERT(journal->j_checkpoint_transactions == NULL);
1169         spin_unlock(&journal->j_list_lock);
1170
1171         if (journal->j_sb_buffer) {
1172                 if (!is_journal_aborted(journal)) {
1173                         /* We can now mark the journal as empty. */
1174                         journal->j_tail = 0;
1175                         journal->j_tail_sequence =
1176                                 ++journal->j_transaction_sequence;
1177                         journal_update_superblock(journal, 1);
1178                 } else {
1179                         err = -EIO;
1180                 }
1181                 brelse(journal->j_sb_buffer);
1182         }
1183
1184         if (journal->j_inode)
1185                 iput(journal->j_inode);
1186         if (journal->j_revoke)
1187                 journal_destroy_revoke(journal);
1188         kfree(journal->j_wbuf);
1189         kfree(journal);
1190
1191         return err;
1192 }
1193
1194
1195 /**
1196  *int journal_check_used_features () - Check if features specified are used.
1197  * @journal: Journal to check.
1198  * @compat: bitmask of compatible features
1199  * @ro: bitmask of features that force read-only mount
1200  * @incompat: bitmask of incompatible features
1201  *
1202  * Check whether the journal uses all of a given set of
1203  * features.  Return true (non-zero) if it does.
1204  **/
1205
1206 int journal_check_used_features (journal_t *journal, unsigned long compat,
1207                                  unsigned long ro, unsigned long incompat)
1208 {
1209         journal_superblock_t *sb;
1210
1211         if (!compat && !ro && !incompat)
1212                 return 1;
1213         if (journal->j_format_version == 1)
1214                 return 0;
1215
1216         sb = journal->j_superblock;
1217
1218         if (((be32_to_cpu(sb->s_feature_compat) & compat) == compat) &&
1219             ((be32_to_cpu(sb->s_feature_ro_compat) & ro) == ro) &&
1220             ((be32_to_cpu(sb->s_feature_incompat) & incompat) == incompat))
1221                 return 1;
1222
1223         return 0;
1224 }
1225
1226 /**
1227  * int journal_check_available_features() - Check feature set in journalling layer
1228  * @journal: Journal to check.
1229  * @compat: bitmask of compatible features
1230  * @ro: bitmask of features that force read-only mount
1231  * @incompat: bitmask of incompatible features
1232  *
1233  * Check whether the journaling code supports the use of
1234  * all of a given set of features on this journal.  Return true
1235  * (non-zero) if it can. */
1236
1237 int journal_check_available_features (journal_t *journal, unsigned long compat,
1238                                       unsigned long ro, unsigned long incompat)
1239 {
1240         journal_superblock_t *sb;
1241
1242         if (!compat && !ro && !incompat)
1243                 return 1;
1244
1245         sb = journal->j_superblock;
1246
1247         /* We can support any known requested features iff the
1248          * superblock is in version 2.  Otherwise we fail to support any
1249          * extended sb features. */
1250
1251         if (journal->j_format_version != 2)
1252                 return 0;
1253
1254         if ((compat   & JFS_KNOWN_COMPAT_FEATURES) == compat &&
1255             (ro       & JFS_KNOWN_ROCOMPAT_FEATURES) == ro &&
1256             (incompat & JFS_KNOWN_INCOMPAT_FEATURES) == incompat)
1257                 return 1;
1258
1259         return 0;
1260 }
1261
1262 /**
1263  * int journal_set_features () - Mark a given journal feature in the superblock
1264  * @journal: Journal to act on.
1265  * @compat: bitmask of compatible features
1266  * @ro: bitmask of features that force read-only mount
1267  * @incompat: bitmask of incompatible features
1268  *
1269  * Mark a given journal feature as present on the
1270  * superblock.  Returns true if the requested features could be set.
1271  *
1272  */
1273
1274 int journal_set_features (journal_t *journal, unsigned long compat,
1275                           unsigned long ro, unsigned long incompat)
1276 {
1277         journal_superblock_t *sb;
1278
1279         if (journal_check_used_features(journal, compat, ro, incompat))
1280                 return 1;
1281
1282         if (!journal_check_available_features(journal, compat, ro, incompat))
1283                 return 0;
1284
1285         jbd_debug(1, "Setting new features 0x%lx/0x%lx/0x%lx\n",
1286                   compat, ro, incompat);
1287
1288         sb = journal->j_superblock;
1289
1290         sb->s_feature_compat    |= cpu_to_be32(compat);
1291         sb->s_feature_ro_compat |= cpu_to_be32(ro);
1292         sb->s_feature_incompat  |= cpu_to_be32(incompat);
1293
1294         return 1;
1295 }
1296
1297
1298 /**
1299  * int journal_update_format () - Update on-disk journal structure.
1300  * @journal: Journal to act on.
1301  *
1302  * Given an initialised but unloaded journal struct, poke about in the
1303  * on-disk structure to update it to the most recent supported version.
1304  */
1305 int journal_update_format (journal_t *journal)
1306 {
1307         journal_superblock_t *sb;
1308         int err;
1309
1310         err = journal_get_superblock(journal);
1311         if (err)
1312                 return err;
1313
1314         sb = journal->j_superblock;
1315
1316         switch (be32_to_cpu(sb->s_header.h_blocktype)) {
1317         case JFS_SUPERBLOCK_V2:
1318                 return 0;
1319         case JFS_SUPERBLOCK_V1:
1320                 return journal_convert_superblock_v1(journal, sb);
1321         default:
1322                 break;
1323         }
1324         return -EINVAL;
1325 }
1326
1327 static int journal_convert_superblock_v1(journal_t *journal,
1328                                          journal_superblock_t *sb)
1329 {
1330         int offset, blocksize;
1331         struct buffer_head *bh;
1332
1333         printk(KERN_WARNING
1334                 "JBD: Converting superblock from version 1 to 2.\n");
1335
1336         /* Pre-initialise new fields to zero */
1337         offset = ((char *) &(sb->s_feature_compat)) - ((char *) sb);
1338         blocksize = be32_to_cpu(sb->s_blocksize);
1339         memset(&sb->s_feature_compat, 0, blocksize-offset);
1340
1341         sb->s_nr_users = cpu_to_be32(1);
1342         sb->s_header.h_blocktype = cpu_to_be32(JFS_SUPERBLOCK_V2);
1343         journal->j_format_version = 2;
1344
1345         bh = journal->j_sb_buffer;
1346         BUFFER_TRACE(bh, "marking dirty");
1347         mark_buffer_dirty(bh);
1348         sync_dirty_buffer(bh);
1349         return 0;
1350 }
1351
1352
1353 /**
1354  * int journal_flush () - Flush journal
1355  * @journal: Journal to act on.
1356  *
1357  * Flush all data for a given journal to disk and empty the journal.
1358  * Filesystems can use this when remounting readonly to ensure that
1359  * recovery does not need to happen on remount.
1360  */
1361
1362 int journal_flush(journal_t *journal)
1363 {
1364         int err = 0;
1365         transaction_t *transaction = NULL;
1366         unsigned long old_tail;
1367
1368         spin_lock(&journal->j_state_lock);
1369
1370         /* Force everything buffered to the log... */
1371         if (journal->j_running_transaction) {
1372                 transaction = journal->j_running_transaction;
1373                 __log_start_commit(journal, transaction->t_tid);
1374         } else if (journal->j_committing_transaction)
1375                 transaction = journal->j_committing_transaction;
1376
1377         /* Wait for the log commit to complete... */
1378         if (transaction) {
1379                 tid_t tid = transaction->t_tid;
1380
1381                 spin_unlock(&journal->j_state_lock);
1382                 log_wait_commit(journal, tid);
1383         } else {
1384                 spin_unlock(&journal->j_state_lock);
1385         }
1386
1387         /* ...and flush everything in the log out to disk. */
1388         spin_lock(&journal->j_list_lock);
1389         while (!err && journal->j_checkpoint_transactions != NULL) {
1390                 spin_unlock(&journal->j_list_lock);
1391                 mutex_lock(&journal->j_checkpoint_mutex);
1392                 err = log_do_checkpoint(journal);
1393                 mutex_unlock(&journal->j_checkpoint_mutex);
1394                 spin_lock(&journal->j_list_lock);
1395         }
1396         spin_unlock(&journal->j_list_lock);
1397
1398         if (is_journal_aborted(journal))
1399                 return -EIO;
1400
1401         cleanup_journal_tail(journal);
1402
1403         /* Finally, mark the journal as really needing no recovery.
1404          * This sets s_start==0 in the underlying superblock, which is
1405          * the magic code for a fully-recovered superblock.  Any future
1406          * commits of data to the journal will restore the current
1407          * s_start value. */
1408         spin_lock(&journal->j_state_lock);
1409         old_tail = journal->j_tail;
1410         journal->j_tail = 0;
1411         spin_unlock(&journal->j_state_lock);
1412         journal_update_superblock(journal, 1);
1413         spin_lock(&journal->j_state_lock);
1414         journal->j_tail = old_tail;
1415
1416         J_ASSERT(!journal->j_running_transaction);
1417         J_ASSERT(!journal->j_committing_transaction);
1418         J_ASSERT(!journal->j_checkpoint_transactions);
1419         J_ASSERT(journal->j_head == journal->j_tail);
1420         J_ASSERT(journal->j_tail_sequence == journal->j_transaction_sequence);
1421         spin_unlock(&journal->j_state_lock);
1422         return 0;
1423 }
1424
1425 /**
1426  * int journal_wipe() - Wipe journal contents
1427  * @journal: Journal to act on.
1428  * @write: flag (see below)
1429  *
1430  * Wipe out all of the contents of a journal, safely.  This will produce
1431  * a warning if the journal contains any valid recovery information.
1432  * Must be called between journal_init_*() and journal_load().
1433  *
1434  * If 'write' is non-zero, then we wipe out the journal on disk; otherwise
1435  * we merely suppress recovery.
1436  */
1437
1438 int journal_wipe(journal_t *journal, int write)
1439 {
1440         journal_superblock_t *sb;
1441         int err = 0;
1442
1443         J_ASSERT (!(journal->j_flags & JFS_LOADED));
1444
1445         err = load_superblock(journal);
1446         if (err)
1447                 return err;
1448
1449         sb = journal->j_superblock;
1450
1451         if (!journal->j_tail)
1452                 goto no_recovery;
1453
1454         printk (KERN_WARNING "JBD: %s recovery information on journal\n",
1455                 write ? "Clearing" : "Ignoring");
1456
1457         err = journal_skip_recovery(journal);
1458         if (write)
1459                 journal_update_superblock(journal, 1);
1460
1461  no_recovery:
1462         return err;
1463 }
1464
1465 /*
1466  * journal_dev_name: format a character string to describe on what
1467  * device this journal is present.
1468  */
1469
1470 static const char *journal_dev_name(journal_t *journal, char *buffer)
1471 {
1472         struct block_device *bdev;
1473
1474         if (journal->j_inode)
1475                 bdev = journal->j_inode->i_sb->s_bdev;
1476         else
1477                 bdev = journal->j_dev;
1478
1479         return bdevname(bdev, buffer);
1480 }
1481
1482 /*
1483  * Journal abort has very specific semantics, which we describe
1484  * for journal abort.
1485  *
1486  * Two internal function, which provide abort to te jbd layer
1487  * itself are here.
1488  */
1489
1490 /*
1491  * Quick version for internal journal use (doesn't lock the journal).
1492  * Aborts hard --- we mark the abort as occurred, but do _nothing_ else,
1493  * and don't attempt to make any other journal updates.
1494  */
1495 static void __journal_abort_hard(journal_t *journal)
1496 {
1497         transaction_t *transaction;
1498         char b[BDEVNAME_SIZE];
1499
1500         if (journal->j_flags & JFS_ABORT)
1501                 return;
1502
1503         printk(KERN_ERR "Aborting journal on device %s.\n",
1504                 journal_dev_name(journal, b));
1505
1506         spin_lock(&journal->j_state_lock);
1507         journal->j_flags |= JFS_ABORT;
1508         transaction = journal->j_running_transaction;
1509         if (transaction)
1510                 __log_start_commit(journal, transaction->t_tid);
1511         spin_unlock(&journal->j_state_lock);
1512 }
1513
1514 /* Soft abort: record the abort error status in the journal superblock,
1515  * but don't do any other IO. */
1516 static void __journal_abort_soft (journal_t *journal, int errno)
1517 {
1518         if (journal->j_flags & JFS_ABORT)
1519                 return;
1520
1521         if (!journal->j_errno)
1522                 journal->j_errno = errno;
1523
1524         __journal_abort_hard(journal);
1525
1526         if (errno)
1527                 journal_update_superblock(journal, 1);
1528 }
1529
1530 /**
1531  * void journal_abort () - Shutdown the journal immediately.
1532  * @journal: the journal to shutdown.
1533  * @errno:   an error number to record in the journal indicating
1534  *           the reason for the shutdown.
1535  *
1536  * Perform a complete, immediate shutdown of the ENTIRE
1537  * journal (not of a single transaction).  This operation cannot be
1538  * undone without closing and reopening the journal.
1539  *
1540  * The journal_abort function is intended to support higher level error
1541  * recovery mechanisms such as the ext2/ext3 remount-readonly error
1542  * mode.
1543  *
1544  * Journal abort has very specific semantics.  Any existing dirty,
1545  * unjournaled buffers in the main filesystem will still be written to
1546  * disk by bdflush, but the journaling mechanism will be suspended
1547  * immediately and no further transaction commits will be honoured.
1548  *
1549  * Any dirty, journaled buffers will be written back to disk without
1550  * hitting the journal.  Atomicity cannot be guaranteed on an aborted
1551  * filesystem, but we _do_ attempt to leave as much data as possible
1552  * behind for fsck to use for cleanup.
1553  *
1554  * Any attempt to get a new transaction handle on a journal which is in
1555  * ABORT state will just result in an -EROFS error return.  A
1556  * journal_stop on an existing handle will return -EIO if we have
1557  * entered abort state during the update.
1558  *
1559  * Recursive transactions are not disturbed by journal abort until the
1560  * final journal_stop, which will receive the -EIO error.
1561  *
1562  * Finally, the journal_abort call allows the caller to supply an errno
1563  * which will be recorded (if possible) in the journal superblock.  This
1564  * allows a client to record failure conditions in the middle of a
1565  * transaction without having to complete the transaction to record the
1566  * failure to disk.  ext3_error, for example, now uses this
1567  * functionality.
1568  *
1569  * Errors which originate from within the journaling layer will NOT
1570  * supply an errno; a null errno implies that absolutely no further
1571  * writes are done to the journal (unless there are any already in
1572  * progress).
1573  *
1574  */
1575
1576 void journal_abort(journal_t *journal, int errno)
1577 {
1578         __journal_abort_soft(journal, errno);
1579 }
1580
1581 /**
1582  * int journal_errno () - returns the journal's error state.
1583  * @journal: journal to examine.
1584  *
1585  * This is the errno numbet set with journal_abort(), the last
1586  * time the journal was mounted - if the journal was stopped
1587  * without calling abort this will be 0.
1588  *
1589  * If the journal has been aborted on this mount time -EROFS will
1590  * be returned.
1591  */
1592 int journal_errno(journal_t *journal)
1593 {
1594         int err;
1595
1596         spin_lock(&journal->j_state_lock);
1597         if (journal->j_flags & JFS_ABORT)
1598                 err = -EROFS;
1599         else
1600                 err = journal->j_errno;
1601         spin_unlock(&journal->j_state_lock);
1602         return err;
1603 }
1604
1605 /**
1606  * int journal_clear_err () - clears the journal's error state
1607  * @journal: journal to act on.
1608  *
1609  * An error must be cleared or Acked to take a FS out of readonly
1610  * mode.
1611  */
1612 int journal_clear_err(journal_t *journal)
1613 {
1614         int err = 0;
1615
1616         spin_lock(&journal->j_state_lock);
1617         if (journal->j_flags & JFS_ABORT)
1618                 err = -EROFS;
1619         else
1620                 journal->j_errno = 0;
1621         spin_unlock(&journal->j_state_lock);
1622         return err;
1623 }
1624
1625 /**
1626  * void journal_ack_err() - Ack journal err.
1627  * @journal: journal to act on.
1628  *
1629  * An error must be cleared or Acked to take a FS out of readonly
1630  * mode.
1631  */
1632 void journal_ack_err(journal_t *journal)
1633 {
1634         spin_lock(&journal->j_state_lock);
1635         if (journal->j_errno)
1636                 journal->j_flags |= JFS_ACK_ERR;
1637         spin_unlock(&journal->j_state_lock);
1638 }
1639
1640 int journal_blocks_per_page(struct inode *inode)
1641 {
1642         return 1 << (PAGE_CACHE_SHIFT - inode->i_sb->s_blocksize_bits);
1643 }
1644
1645 /*
1646  * Journal_head storage management
1647  */
1648 static struct kmem_cache *journal_head_cache;
1649 #ifdef CONFIG_JBD_DEBUG
1650 static atomic_t nr_journal_heads = ATOMIC_INIT(0);
1651 #endif
1652
1653 static int journal_init_journal_head_cache(void)
1654 {
1655         int retval;
1656
1657         J_ASSERT(journal_head_cache == NULL);
1658         journal_head_cache = kmem_cache_create("journal_head",
1659                                 sizeof(struct journal_head),
1660                                 0,              /* offset */
1661                                 SLAB_TEMPORARY, /* flags */
1662                                 NULL);          /* ctor */
1663         retval = 0;
1664         if (!journal_head_cache) {
1665                 retval = -ENOMEM;
1666                 printk(KERN_EMERG "JBD: no memory for journal_head cache\n");
1667         }
1668         return retval;
1669 }
1670
1671 static void journal_destroy_journal_head_cache(void)
1672 {
1673         if (journal_head_cache) {
1674                 kmem_cache_destroy(journal_head_cache);
1675                 journal_head_cache = NULL;
1676         }
1677 }
1678
1679 /*
1680  * journal_head splicing and dicing
1681  */
1682 static struct journal_head *journal_alloc_journal_head(void)
1683 {
1684         struct journal_head *ret;
1685         static unsigned long last_warning;
1686
1687 #ifdef CONFIG_JBD_DEBUG
1688         atomic_inc(&nr_journal_heads);
1689 #endif
1690         ret = kmem_cache_alloc(journal_head_cache, GFP_NOFS);
1691         if (ret == NULL) {
1692                 jbd_debug(1, "out of memory for journal_head\n");
1693                 if (time_after(jiffies, last_warning + 5*HZ)) {
1694                         printk(KERN_NOTICE "ENOMEM in %s, retrying.\n",
1695                                __func__);
1696                         last_warning = jiffies;
1697                 }
1698                 while (ret == NULL) {
1699                         yield();
1700                         ret = kmem_cache_alloc(journal_head_cache, GFP_NOFS);
1701                 }
1702         }
1703         return ret;
1704 }
1705
1706 static void journal_free_journal_head(struct journal_head *jh)
1707 {
1708 #ifdef CONFIG_JBD_DEBUG
1709         atomic_dec(&nr_journal_heads);
1710         memset(jh, JBD_POISON_FREE, sizeof(*jh));
1711 #endif
1712         kmem_cache_free(journal_head_cache, jh);
1713 }
1714
1715 /*
1716  * A journal_head is attached to a buffer_head whenever JBD has an
1717  * interest in the buffer.
1718  *
1719  * Whenever a buffer has an attached journal_head, its ->b_state:BH_JBD bit
1720  * is set.  This bit is tested in core kernel code where we need to take
1721  * JBD-specific actions.  Testing the zeroness of ->b_private is not reliable
1722  * there.
1723  *
1724  * When a buffer has its BH_JBD bit set, its ->b_count is elevated by one.
1725  *
1726  * When a buffer has its BH_JBD bit set it is immune from being released by
1727  * core kernel code, mainly via ->b_count.
1728  *
1729  * A journal_head may be detached from its buffer_head when the journal_head's
1730  * b_transaction, b_cp_transaction and b_next_transaction pointers are NULL.
1731  * Various places in JBD call journal_remove_journal_head() to indicate that the
1732  * journal_head can be dropped if needed.
1733  *
1734  * Various places in the kernel want to attach a journal_head to a buffer_head
1735  * _before_ attaching the journal_head to a transaction.  To protect the
1736  * journal_head in this situation, journal_add_journal_head elevates the
1737  * journal_head's b_jcount refcount by one.  The caller must call
1738  * journal_put_journal_head() to undo this.
1739  *
1740  * So the typical usage would be:
1741  *
1742  *      (Attach a journal_head if needed.  Increments b_jcount)
1743  *      struct journal_head *jh = journal_add_journal_head(bh);
1744  *      ...
1745  *      jh->b_transaction = xxx;
1746  *      journal_put_journal_head(jh);
1747  *
1748  * Now, the journal_head's b_jcount is zero, but it is safe from being released
1749  * because it has a non-zero b_transaction.
1750  */
1751
1752 /*
1753  * Give a buffer_head a journal_head.
1754  *
1755  * Doesn't need the journal lock.
1756  * May sleep.
1757  */
1758 struct journal_head *journal_add_journal_head(struct buffer_head *bh)
1759 {
1760         struct journal_head *jh;
1761         struct journal_head *new_jh = NULL;
1762
1763 repeat:
1764         if (!buffer_jbd(bh)) {
1765                 new_jh = journal_alloc_journal_head();
1766                 memset(new_jh, 0, sizeof(*new_jh));
1767         }
1768
1769         jbd_lock_bh_journal_head(bh);
1770         if (buffer_jbd(bh)) {
1771                 jh = bh2jh(bh);
1772         } else {
1773                 J_ASSERT_BH(bh,
1774                         (atomic_read(&bh->b_count) > 0) ||
1775                         (bh->b_page && bh->b_page->mapping));
1776
1777                 if (!new_jh) {
1778                         jbd_unlock_bh_journal_head(bh);
1779                         goto repeat;
1780                 }
1781
1782                 jh = new_jh;
1783                 new_jh = NULL;          /* We consumed it */
1784                 set_buffer_jbd(bh);
1785                 bh->b_private = jh;
1786                 jh->b_bh = bh;
1787                 get_bh(bh);
1788                 BUFFER_TRACE(bh, "added journal_head");
1789         }
1790         jh->b_jcount++;
1791         jbd_unlock_bh_journal_head(bh);
1792         if (new_jh)
1793                 journal_free_journal_head(new_jh);
1794         return bh->b_private;
1795 }
1796
1797 /*
1798  * Grab a ref against this buffer_head's journal_head.  If it ended up not
1799  * having a journal_head, return NULL
1800  */
1801 struct journal_head *journal_grab_journal_head(struct buffer_head *bh)
1802 {
1803         struct journal_head *jh = NULL;
1804
1805         jbd_lock_bh_journal_head(bh);
1806         if (buffer_jbd(bh)) {
1807                 jh = bh2jh(bh);
1808                 jh->b_jcount++;
1809         }
1810         jbd_unlock_bh_journal_head(bh);
1811         return jh;
1812 }
1813
1814 static void __journal_remove_journal_head(struct buffer_head *bh)
1815 {
1816         struct journal_head *jh = bh2jh(bh);
1817
1818         J_ASSERT_JH(jh, jh->b_jcount >= 0);
1819
1820         get_bh(bh);
1821         if (jh->b_jcount == 0) {
1822                 if (jh->b_transaction == NULL &&
1823                                 jh->b_next_transaction == NULL &&
1824                                 jh->b_cp_transaction == NULL) {
1825                         J_ASSERT_JH(jh, jh->b_jlist == BJ_None);
1826                         J_ASSERT_BH(bh, buffer_jbd(bh));
1827                         J_ASSERT_BH(bh, jh2bh(jh) == bh);
1828                         BUFFER_TRACE(bh, "remove journal_head");
1829                         if (jh->b_frozen_data) {
1830                                 printk(KERN_WARNING "%s: freeing "
1831                                                 "b_frozen_data\n",
1832                                                 __func__);
1833                                 jbd_free(jh->b_frozen_data, bh->b_size);
1834                         }
1835                         if (jh->b_committed_data) {
1836                                 printk(KERN_WARNING "%s: freeing "
1837                                                 "b_committed_data\n",
1838                                                 __func__);
1839                                 jbd_free(jh->b_committed_data, bh->b_size);
1840                         }
1841                         bh->b_private = NULL;
1842                         jh->b_bh = NULL;        /* debug, really */
1843                         clear_buffer_jbd(bh);
1844                         __brelse(bh);
1845                         journal_free_journal_head(jh);
1846                 } else {
1847                         BUFFER_TRACE(bh, "journal_head was locked");
1848                 }
1849         }
1850 }
1851
1852 /*
1853  * journal_remove_journal_head(): if the buffer isn't attached to a transaction
1854  * and has a zero b_jcount then remove and release its journal_head.   If we did
1855  * see that the buffer is not used by any transaction we also "logically"
1856  * decrement ->b_count.
1857  *
1858  * We in fact take an additional increment on ->b_count as a convenience,
1859  * because the caller usually wants to do additional things with the bh
1860  * after calling here.
1861  * The caller of journal_remove_journal_head() *must* run __brelse(bh) at some
1862  * time.  Once the caller has run __brelse(), the buffer is eligible for
1863  * reaping by try_to_free_buffers().
1864  */
1865 void journal_remove_journal_head(struct buffer_head *bh)
1866 {
1867         jbd_lock_bh_journal_head(bh);
1868         __journal_remove_journal_head(bh);
1869         jbd_unlock_bh_journal_head(bh);
1870 }
1871
1872 /*
1873  * Drop a reference on the passed journal_head.  If it fell to zero then try to
1874  * release the journal_head from the buffer_head.
1875  */
1876 void journal_put_journal_head(struct journal_head *jh)
1877 {
1878         struct buffer_head *bh = jh2bh(jh);
1879
1880         jbd_lock_bh_journal_head(bh);
1881         J_ASSERT_JH(jh, jh->b_jcount > 0);
1882         --jh->b_jcount;
1883         if (!jh->b_jcount && !jh->b_transaction) {
1884                 __journal_remove_journal_head(bh);
1885                 __brelse(bh);
1886         }
1887         jbd_unlock_bh_journal_head(bh);
1888 }
1889
1890 /*
1891  * debugfs tunables
1892  */
1893 #ifdef CONFIG_JBD_DEBUG
1894
1895 u8 journal_enable_debug __read_mostly;
1896 EXPORT_SYMBOL(journal_enable_debug);
1897
1898 static struct dentry *jbd_debugfs_dir;
1899 static struct dentry *jbd_debug;
1900
1901 static void __init jbd_create_debugfs_entry(void)
1902 {
1903         jbd_debugfs_dir = debugfs_create_dir("jbd", NULL);
1904         if (jbd_debugfs_dir)
1905                 jbd_debug = debugfs_create_u8("jbd-debug", S_IRUGO,
1906                                                jbd_debugfs_dir,
1907                                                &journal_enable_debug);
1908 }
1909
1910 static void __exit jbd_remove_debugfs_entry(void)
1911 {
1912         debugfs_remove(jbd_debug);
1913         debugfs_remove(jbd_debugfs_dir);
1914 }
1915
1916 #else
1917
1918 static inline void jbd_create_debugfs_entry(void)
1919 {
1920 }
1921
1922 static inline void jbd_remove_debugfs_entry(void)
1923 {
1924 }
1925
1926 #endif
1927
1928 struct kmem_cache *jbd_handle_cache;
1929
1930 static int __init journal_init_handle_cache(void)
1931 {
1932         jbd_handle_cache = kmem_cache_create("journal_handle",
1933                                 sizeof(handle_t),
1934                                 0,              /* offset */
1935                                 SLAB_TEMPORARY, /* flags */
1936                                 NULL);          /* ctor */
1937         if (jbd_handle_cache == NULL) {
1938                 printk(KERN_EMERG "JBD: failed to create handle cache\n");
1939                 return -ENOMEM;
1940         }
1941         return 0;
1942 }
1943
1944 static void journal_destroy_handle_cache(void)
1945 {
1946         if (jbd_handle_cache)
1947                 kmem_cache_destroy(jbd_handle_cache);
1948 }
1949
1950 /*
1951  * Module startup and shutdown
1952  */
1953
1954 static int __init journal_init_caches(void)
1955 {
1956         int ret;
1957
1958         ret = journal_init_revoke_caches();
1959         if (ret == 0)
1960                 ret = journal_init_journal_head_cache();
1961         if (ret == 0)
1962                 ret = journal_init_handle_cache();
1963         return ret;
1964 }
1965
1966 static void journal_destroy_caches(void)
1967 {
1968         journal_destroy_revoke_caches();
1969         journal_destroy_journal_head_cache();
1970         journal_destroy_handle_cache();
1971 }
1972
1973 static int __init journal_init(void)
1974 {
1975         int ret;
1976
1977         BUILD_BUG_ON(sizeof(struct journal_superblock_s) != 1024);
1978
1979         ret = journal_init_caches();
1980         if (ret != 0)
1981                 journal_destroy_caches();
1982         jbd_create_debugfs_entry();
1983         return ret;
1984 }
1985
1986 static void __exit journal_exit(void)
1987 {
1988 #ifdef CONFIG_JBD_DEBUG
1989         int n = atomic_read(&nr_journal_heads);
1990         if (n)
1991                 printk(KERN_EMERG "JBD: leaked %d journal_heads!\n", n);
1992 #endif
1993         jbd_remove_debugfs_entry();
1994         journal_destroy_caches();
1995 }
1996
1997 MODULE_LICENSE("GPL");
1998 module_init(journal_init);
1999 module_exit(journal_exit);
2000