Merge git://git.kernel.org/pub/scm/linux/kernel/git/davem/sparc-2.6
[linux-2.6.git] / 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(log_start_commit);
77 EXPORT_SYMBOL(journal_start_commit);
78 EXPORT_SYMBOL(journal_force_commit_nested);
79 EXPORT_SYMBOL(journal_wipe);
80 EXPORT_SYMBOL(journal_blocks_per_page);
81 EXPORT_SYMBOL(journal_invalidatepage);
82 EXPORT_SYMBOL(journal_try_to_free_buffers);
83 EXPORT_SYMBOL(journal_force_commit);
84
85 static int journal_convert_superblock_v1(journal_t *, journal_superblock_t *);
86 static void __journal_abort_soft (journal_t *journal, int errno);
87
88 /*
89  * Helper function used to manage commit timeouts
90  */
91
92 static void commit_timeout(unsigned long __data)
93 {
94         struct task_struct * p = (struct task_struct *) __data;
95
96         wake_up_process(p);
97 }
98
99 /*
100  * kjournald: The main thread function used to manage a logging device
101  * journal.
102  *
103  * This kernel thread is responsible for two things:
104  *
105  * 1) COMMIT:  Every so often we need to commit the current state of the
106  *    filesystem to disk.  The journal thread is responsible for writing
107  *    all of the metadata buffers to disk.
108  *
109  * 2) CHECKPOINT: We cannot reuse a used section of the log file until all
110  *    of the data in that part of the log has been rewritten elsewhere on
111  *    the disk.  Flushing these old buffers to reclaim space in the log is
112  *    known as checkpointing, and this thread is responsible for that job.
113  */
114
115 static int kjournald(void *arg)
116 {
117         journal_t *journal = arg;
118         transaction_t *transaction;
119
120         /*
121          * Set up an interval timer which can be used to trigger a commit wakeup
122          * after the commit interval expires
123          */
124         setup_timer(&journal->j_commit_timer, commit_timeout,
125                         (unsigned long)current);
126
127         /* Record that the journal thread is running */
128         journal->j_task = current;
129         wake_up(&journal->j_wait_done_commit);
130
131         printk(KERN_INFO "kjournald starting.  Commit interval %ld seconds\n",
132                         journal->j_commit_interval / HZ);
133
134         /*
135          * And now, wait forever for commit wakeup events.
136          */
137         spin_lock(&journal->j_state_lock);
138
139 loop:
140         if (journal->j_flags & JFS_UNMOUNT)
141                 goto end_loop;
142
143         jbd_debug(1, "commit_sequence=%d, commit_request=%d\n",
144                 journal->j_commit_sequence, journal->j_commit_request);
145
146         if (journal->j_commit_sequence != journal->j_commit_request) {
147                 jbd_debug(1, "OK, requests differ\n");
148                 spin_unlock(&journal->j_state_lock);
149                 del_timer_sync(&journal->j_commit_timer);
150                 journal_commit_transaction(journal);
151                 spin_lock(&journal->j_state_lock);
152                 goto loop;
153         }
154
155         wake_up(&journal->j_wait_done_commit);
156         if (freezing(current)) {
157                 /*
158                  * The simpler the better. Flushing journal isn't a
159                  * good idea, because that depends on threads that may
160                  * be already stopped.
161                  */
162                 jbd_debug(1, "Now suspending kjournald\n");
163                 spin_unlock(&journal->j_state_lock);
164                 refrigerator();
165                 spin_lock(&journal->j_state_lock);
166         } else {
167                 /*
168                  * We assume on resume that commits are already there,
169                  * so we don't sleep
170                  */
171                 DEFINE_WAIT(wait);
172                 int should_sleep = 1;
173
174                 prepare_to_wait(&journal->j_wait_commit, &wait,
175                                 TASK_INTERRUPTIBLE);
176                 if (journal->j_commit_sequence != journal->j_commit_request)
177                         should_sleep = 0;
178                 transaction = journal->j_running_transaction;
179                 if (transaction && time_after_eq(jiffies,
180                                                 transaction->t_expires))
181                         should_sleep = 0;
182                 if (journal->j_flags & JFS_UNMOUNT)
183                         should_sleep = 0;
184                 if (should_sleep) {
185                         spin_unlock(&journal->j_state_lock);
186                         schedule();
187                         spin_lock(&journal->j_state_lock);
188                 }
189                 finish_wait(&journal->j_wait_commit, &wait);
190         }
191
192         jbd_debug(1, "kjournald wakes\n");
193
194         /*
195          * Were we woken up by a commit wakeup event?
196          */
197         transaction = journal->j_running_transaction;
198         if (transaction && time_after_eq(jiffies, transaction->t_expires)) {
199                 journal->j_commit_request = transaction->t_tid;
200                 jbd_debug(1, "woke because of timeout\n");
201         }
202         goto loop;
203
204 end_loop:
205         spin_unlock(&journal->j_state_lock);
206         del_timer_sync(&journal->j_commit_timer);
207         journal->j_task = NULL;
208         wake_up(&journal->j_wait_done_commit);
209         jbd_debug(1, "Journal thread exiting.\n");
210         return 0;
211 }
212
213 static int journal_start_thread(journal_t *journal)
214 {
215         struct task_struct *t;
216
217         t = kthread_run(kjournald, journal, "kjournald");
218         if (IS_ERR(t))
219                 return PTR_ERR(t);
220
221         wait_event(journal->j_wait_done_commit, journal->j_task != NULL);
222         return 0;
223 }
224
225 static void journal_kill_thread(journal_t *journal)
226 {
227         spin_lock(&journal->j_state_lock);
228         journal->j_flags |= JFS_UNMOUNT;
229
230         while (journal->j_task) {
231                 wake_up(&journal->j_wait_commit);
232                 spin_unlock(&journal->j_state_lock);
233                 wait_event(journal->j_wait_done_commit,
234                                 journal->j_task == NULL);
235                 spin_lock(&journal->j_state_lock);
236         }
237         spin_unlock(&journal->j_state_lock);
238 }
239
240 /*
241  * journal_write_metadata_buffer: write a metadata buffer to the journal.
242  *
243  * Writes a metadata buffer to a given disk block.  The actual IO is not
244  * performed but a new buffer_head is constructed which labels the data
245  * to be written with the correct destination disk block.
246  *
247  * Any magic-number escaping which needs to be done will cause a
248  * copy-out here.  If the buffer happens to start with the
249  * JFS_MAGIC_NUMBER, then we can't write it to the log directly: the
250  * magic number is only written to the log for descripter blocks.  In
251  * this case, we copy the data and replace the first word with 0, and we
252  * return a result code which indicates that this buffer needs to be
253  * marked as an escaped buffer in the corresponding log descriptor
254  * block.  The missing word can then be restored when the block is read
255  * during recovery.
256  *
257  * If the source buffer has already been modified by a new transaction
258  * since we took the last commit snapshot, we use the frozen copy of
259  * that data for IO.  If we end up using the existing buffer_head's data
260  * for the write, then we *have* to lock the buffer to prevent anyone
261  * else from using and possibly modifying it while the IO is in
262  * progress.
263  *
264  * The function returns a pointer to the buffer_heads to be used for IO.
265  *
266  * We assume that the journal has already been locked in this function.
267  *
268  * Return value:
269  *  <0: Error
270  * >=0: Finished OK
271  *
272  * On success:
273  * Bit 0 set == escape performed on the data
274  * Bit 1 set == buffer copy-out performed (kfree the data after IO)
275  */
276
277 int journal_write_metadata_buffer(transaction_t *transaction,
278                                   struct journal_head  *jh_in,
279                                   struct journal_head **jh_out,
280                                   unsigned int blocknr)
281 {
282         int need_copy_out = 0;
283         int done_copy_out = 0;
284         int do_escape = 0;
285         char *mapped_data;
286         struct buffer_head *new_bh;
287         struct journal_head *new_jh;
288         struct page *new_page;
289         unsigned int new_offset;
290         struct buffer_head *bh_in = jh2bh(jh_in);
291         journal_t *journal = transaction->t_journal;
292
293         /*
294          * The buffer really shouldn't be locked: only the current committing
295          * transaction is allowed to write it, so nobody else is allowed
296          * to do any IO.
297          *
298          * akpm: except if we're journalling data, and write() output is
299          * also part of a shared mapping, and another thread has
300          * decided to launch a writepage() against this buffer.
301          */
302         J_ASSERT_BH(bh_in, buffer_jbddirty(bh_in));
303
304         new_bh = alloc_buffer_head(GFP_NOFS|__GFP_NOFAIL);
305         /* keep subsequent assertions sane */
306         new_bh->b_state = 0;
307         init_buffer(new_bh, NULL, NULL);
308         atomic_set(&new_bh->b_count, 1);
309         new_jh = journal_add_journal_head(new_bh);      /* This sleeps */
310
311         /*
312          * If a new transaction has already done a buffer copy-out, then
313          * we use that version of the data for the commit.
314          */
315         jbd_lock_bh_state(bh_in);
316 repeat:
317         if (jh_in->b_frozen_data) {
318                 done_copy_out = 1;
319                 new_page = virt_to_page(jh_in->b_frozen_data);
320                 new_offset = offset_in_page(jh_in->b_frozen_data);
321         } else {
322                 new_page = jh2bh(jh_in)->b_page;
323                 new_offset = offset_in_page(jh2bh(jh_in)->b_data);
324         }
325
326         mapped_data = kmap_atomic(new_page, KM_USER0);
327         /*
328          * Check for escaping
329          */
330         if (*((__be32 *)(mapped_data + new_offset)) ==
331                                 cpu_to_be32(JFS_MAGIC_NUMBER)) {
332                 need_copy_out = 1;
333                 do_escape = 1;
334         }
335         kunmap_atomic(mapped_data, KM_USER0);
336
337         /*
338          * Do we need to do a data copy?
339          */
340         if (need_copy_out && !done_copy_out) {
341                 char *tmp;
342
343                 jbd_unlock_bh_state(bh_in);
344                 tmp = jbd_alloc(bh_in->b_size, GFP_NOFS);
345                 jbd_lock_bh_state(bh_in);
346                 if (jh_in->b_frozen_data) {
347                         jbd_free(tmp, bh_in->b_size);
348                         goto repeat;
349                 }
350
351                 jh_in->b_frozen_data = tmp;
352                 mapped_data = kmap_atomic(new_page, KM_USER0);
353                 memcpy(tmp, mapped_data + new_offset, jh2bh(jh_in)->b_size);
354                 kunmap_atomic(mapped_data, KM_USER0);
355
356                 new_page = virt_to_page(tmp);
357                 new_offset = offset_in_page(tmp);
358                 done_copy_out = 1;
359         }
360
361         /*
362          * Did we need to do an escaping?  Now we've done all the
363          * copying, we can finally do so.
364          */
365         if (do_escape) {
366                 mapped_data = kmap_atomic(new_page, KM_USER0);
367                 *((unsigned int *)(mapped_data + new_offset)) = 0;
368                 kunmap_atomic(mapped_data, KM_USER0);
369         }
370
371         set_bh_page(new_bh, new_page, new_offset);
372         new_jh->b_transaction = NULL;
373         new_bh->b_size = jh2bh(jh_in)->b_size;
374         new_bh->b_bdev = transaction->t_journal->j_dev;
375         new_bh->b_blocknr = blocknr;
376         set_buffer_mapped(new_bh);
377         set_buffer_dirty(new_bh);
378
379         *jh_out = new_jh;
380
381         /*
382          * The to-be-written buffer needs to get moved to the io queue,
383          * and the original buffer whose contents we are shadowing or
384          * copying is moved to the transaction's shadow queue.
385          */
386         JBUFFER_TRACE(jh_in, "file as BJ_Shadow");
387         spin_lock(&journal->j_list_lock);
388         __journal_file_buffer(jh_in, transaction, BJ_Shadow);
389         spin_unlock(&journal->j_list_lock);
390         jbd_unlock_bh_state(bh_in);
391
392         JBUFFER_TRACE(new_jh, "file as BJ_IO");
393         journal_file_buffer(new_jh, transaction, BJ_IO);
394
395         return do_escape | (done_copy_out << 1);
396 }
397
398 /*
399  * Allocation code for the journal file.  Manage the space left in the
400  * journal, so that we can begin checkpointing when appropriate.
401  */
402
403 /*
404  * __log_space_left: Return the number of free blocks left in the journal.
405  *
406  * Called with the journal already locked.
407  *
408  * Called under j_state_lock
409  */
410
411 int __log_space_left(journal_t *journal)
412 {
413         int left = journal->j_free;
414
415         assert_spin_locked(&journal->j_state_lock);
416
417         /*
418          * Be pessimistic here about the number of those free blocks which
419          * might be required for log descriptor control blocks.
420          */
421
422 #define MIN_LOG_RESERVED_BLOCKS 32 /* Allow for rounding errors */
423
424         left -= MIN_LOG_RESERVED_BLOCKS;
425
426         if (left <= 0)
427                 return 0;
428         left -= (left >> 3);
429         return left;
430 }
431
432 /*
433  * Called under j_state_lock.  Returns true if a transaction commit was started.
434  */
435 int __log_start_commit(journal_t *journal, tid_t target)
436 {
437         /*
438          * Are we already doing a recent enough commit?
439          */
440         if (!tid_geq(journal->j_commit_request, target)) {
441                 /*
442                  * We want a new commit: OK, mark the request and wakup the
443                  * commit thread.  We do _not_ do the commit ourselves.
444                  */
445
446                 journal->j_commit_request = target;
447                 jbd_debug(1, "JBD: requesting commit %d/%d\n",
448                           journal->j_commit_request,
449                           journal->j_commit_sequence);
450                 wake_up(&journal->j_wait_commit);
451                 return 1;
452         }
453         return 0;
454 }
455
456 int log_start_commit(journal_t *journal, tid_t tid)
457 {
458         int ret;
459
460         spin_lock(&journal->j_state_lock);
461         ret = __log_start_commit(journal, tid);
462         spin_unlock(&journal->j_state_lock);
463         return ret;
464 }
465
466 /*
467  * Force and wait upon a commit if the calling process is not within
468  * transaction.  This is used for forcing out undo-protected data which contains
469  * bitmaps, when the fs is running out of space.
470  *
471  * We can only force the running transaction if we don't have an active handle;
472  * otherwise, we will deadlock.
473  *
474  * Returns true if a transaction was started.
475  */
476 int journal_force_commit_nested(journal_t *journal)
477 {
478         transaction_t *transaction = NULL;
479         tid_t tid;
480
481         spin_lock(&journal->j_state_lock);
482         if (journal->j_running_transaction && !current->journal_info) {
483                 transaction = journal->j_running_transaction;
484                 __log_start_commit(journal, transaction->t_tid);
485         } else if (journal->j_committing_transaction)
486                 transaction = journal->j_committing_transaction;
487
488         if (!transaction) {
489                 spin_unlock(&journal->j_state_lock);
490                 return 0;       /* Nothing to retry */
491         }
492
493         tid = transaction->t_tid;
494         spin_unlock(&journal->j_state_lock);
495         log_wait_commit(journal, tid);
496         return 1;
497 }
498
499 /*
500  * Start a commit of the current running transaction (if any).  Returns true
501  * if a transaction is going to be committed (or is currently already
502  * committing), and fills its tid in at *ptid
503  */
504 int journal_start_commit(journal_t *journal, tid_t *ptid)
505 {
506         int ret = 0;
507
508         spin_lock(&journal->j_state_lock);
509         if (journal->j_running_transaction) {
510                 tid_t tid = journal->j_running_transaction->t_tid;
511
512                 __log_start_commit(journal, tid);
513                 /* There's a running transaction and we've just made sure
514                  * it's commit has been scheduled. */
515                 if (ptid)
516                         *ptid = tid;
517                 ret = 1;
518         } else if (journal->j_committing_transaction) {
519                 /*
520                  * If ext3_write_super() recently started a commit, then we
521                  * have to wait for completion of that transaction
522                  */
523                 if (ptid)
524                         *ptid = journal->j_committing_transaction->t_tid;
525                 ret = 1;
526         }
527         spin_unlock(&journal->j_state_lock);
528         return ret;
529 }
530
531 /*
532  * Wait for a specified commit to complete.
533  * The caller may not hold the journal lock.
534  */
535 int log_wait_commit(journal_t *journal, tid_t tid)
536 {
537         int err = 0;
538
539 #ifdef CONFIG_JBD_DEBUG
540         spin_lock(&journal->j_state_lock);
541         if (!tid_geq(journal->j_commit_request, tid)) {
542                 printk(KERN_EMERG
543                        "%s: error: j_commit_request=%d, tid=%d\n",
544                        __func__, journal->j_commit_request, tid);
545         }
546         spin_unlock(&journal->j_state_lock);
547 #endif
548         spin_lock(&journal->j_state_lock);
549         while (tid_gt(tid, journal->j_commit_sequence)) {
550                 jbd_debug(1, "JBD: want %d, j_commit_sequence=%d\n",
551                                   tid, journal->j_commit_sequence);
552                 wake_up(&journal->j_wait_commit);
553                 spin_unlock(&journal->j_state_lock);
554                 wait_event(journal->j_wait_done_commit,
555                                 !tid_gt(tid, journal->j_commit_sequence));
556                 spin_lock(&journal->j_state_lock);
557         }
558         spin_unlock(&journal->j_state_lock);
559
560         if (unlikely(is_journal_aborted(journal))) {
561                 printk(KERN_EMERG "journal commit I/O error\n");
562                 err = -EIO;
563         }
564         return err;
565 }
566
567 /*
568  * Log buffer allocation routines:
569  */
570
571 int journal_next_log_block(journal_t *journal, unsigned int *retp)
572 {
573         unsigned int blocknr;
574
575         spin_lock(&journal->j_state_lock);
576         J_ASSERT(journal->j_free > 1);
577
578         blocknr = journal->j_head;
579         journal->j_head++;
580         journal->j_free--;
581         if (journal->j_head == journal->j_last)
582                 journal->j_head = journal->j_first;
583         spin_unlock(&journal->j_state_lock);
584         return journal_bmap(journal, blocknr, retp);
585 }
586
587 /*
588  * Conversion of logical to physical block numbers for the journal
589  *
590  * On external journals the journal blocks are identity-mapped, so
591  * this is a no-op.  If needed, we can use j_blk_offset - everything is
592  * ready.
593  */
594 int journal_bmap(journal_t *journal, unsigned int blocknr,
595                  unsigned int *retp)
596 {
597         int err = 0;
598         unsigned int ret;
599
600         if (journal->j_inode) {
601                 ret = bmap(journal->j_inode, blocknr);
602                 if (ret)
603                         *retp = ret;
604                 else {
605                         char b[BDEVNAME_SIZE];
606
607                         printk(KERN_ALERT "%s: journal block not found "
608                                         "at offset %u on %s\n",
609                                 __func__,
610                                 blocknr,
611                                 bdevname(journal->j_dev, b));
612                         err = -EIO;
613                         __journal_abort_soft(journal, err);
614                 }
615         } else {
616                 *retp = blocknr; /* +journal->j_blk_offset */
617         }
618         return err;
619 }
620
621 /*
622  * We play buffer_head aliasing tricks to write data/metadata blocks to
623  * the journal without copying their contents, but for journal
624  * descriptor blocks we do need to generate bona fide buffers.
625  *
626  * After the caller of journal_get_descriptor_buffer() has finished modifying
627  * the buffer's contents they really should run flush_dcache_page(bh->b_page).
628  * But we don't bother doing that, so there will be coherency problems with
629  * mmaps of blockdevs which hold live JBD-controlled filesystems.
630  */
631 struct journal_head *journal_get_descriptor_buffer(journal_t *journal)
632 {
633         struct buffer_head *bh;
634         unsigned int blocknr;
635         int err;
636
637         err = journal_next_log_block(journal, &blocknr);
638
639         if (err)
640                 return NULL;
641
642         bh = __getblk(journal->j_dev, blocknr, journal->j_blocksize);
643         if (!bh)
644                 return NULL;
645         lock_buffer(bh);
646         memset(bh->b_data, 0, journal->j_blocksize);
647         set_buffer_uptodate(bh);
648         unlock_buffer(bh);
649         BUFFER_TRACE(bh, "return this buffer");
650         return journal_add_journal_head(bh);
651 }
652
653 /*
654  * Management for journal control blocks: functions to create and
655  * destroy journal_t structures, and to initialise and read existing
656  * journal blocks from disk.  */
657
658 /* First: create and setup a journal_t object in memory.  We initialise
659  * very few fields yet: that has to wait until we have created the
660  * journal structures from from scratch, or loaded them from disk. */
661
662 static journal_t * journal_init_common (void)
663 {
664         journal_t *journal;
665         int err;
666
667         journal = kzalloc(sizeof(*journal), GFP_KERNEL);
668         if (!journal)
669                 goto fail;
670
671         init_waitqueue_head(&journal->j_wait_transaction_locked);
672         init_waitqueue_head(&journal->j_wait_logspace);
673         init_waitqueue_head(&journal->j_wait_done_commit);
674         init_waitqueue_head(&journal->j_wait_checkpoint);
675         init_waitqueue_head(&journal->j_wait_commit);
676         init_waitqueue_head(&journal->j_wait_updates);
677         mutex_init(&journal->j_barrier);
678         mutex_init(&journal->j_checkpoint_mutex);
679         spin_lock_init(&journal->j_revoke_lock);
680         spin_lock_init(&journal->j_list_lock);
681         spin_lock_init(&journal->j_state_lock);
682
683         journal->j_commit_interval = (HZ * JBD_DEFAULT_MAX_COMMIT_AGE);
684
685         /* The journal is marked for error until we succeed with recovery! */
686         journal->j_flags = JFS_ABORT;
687
688         /* Set up a default-sized revoke table for the new mount. */
689         err = journal_init_revoke(journal, JOURNAL_REVOKE_DEFAULT_HASH);
690         if (err) {
691                 kfree(journal);
692                 goto fail;
693         }
694         return journal;
695 fail:
696         return NULL;
697 }
698
699 /* journal_init_dev and journal_init_inode:
700  *
701  * Create a journal structure assigned some fixed set of disk blocks to
702  * the journal.  We don't actually touch those disk blocks yet, but we
703  * need to set up all of the mapping information to tell the journaling
704  * system where the journal blocks are.
705  *
706  */
707
708 /**
709  *  journal_t * journal_init_dev() - creates and initialises a journal structure
710  *  @bdev: Block device on which to create the journal
711  *  @fs_dev: Device which hold journalled filesystem for this journal.
712  *  @start: Block nr Start of journal.
713  *  @len:  Length of the journal in blocks.
714  *  @blocksize: blocksize of journalling device
715  *
716  *  Returns: a newly created journal_t *
717  *
718  *  journal_init_dev creates a journal which maps a fixed contiguous
719  *  range of blocks on an arbitrary block device.
720  *
721  */
722 journal_t * journal_init_dev(struct block_device *bdev,
723                         struct block_device *fs_dev,
724                         int start, int len, int blocksize)
725 {
726         journal_t *journal = journal_init_common();
727         struct buffer_head *bh;
728         int n;
729
730         if (!journal)
731                 return NULL;
732
733         /* journal descriptor can store up to n blocks -bzzz */
734         journal->j_blocksize = blocksize;
735         n = journal->j_blocksize / sizeof(journal_block_tag_t);
736         journal->j_wbufsize = n;
737         journal->j_wbuf = kmalloc(n * sizeof(struct buffer_head*), GFP_KERNEL);
738         if (!journal->j_wbuf) {
739                 printk(KERN_ERR "%s: Cant allocate bhs for commit thread\n",
740                         __func__);
741                 goto out_err;
742         }
743         journal->j_dev = bdev;
744         journal->j_fs_dev = fs_dev;
745         journal->j_blk_offset = start;
746         journal->j_maxlen = len;
747
748         bh = __getblk(journal->j_dev, start, journal->j_blocksize);
749         if (!bh) {
750                 printk(KERN_ERR
751                        "%s: Cannot get buffer for journal superblock\n",
752                        __func__);
753                 goto out_err;
754         }
755         journal->j_sb_buffer = bh;
756         journal->j_superblock = (journal_superblock_t *)bh->b_data;
757
758         return journal;
759 out_err:
760         kfree(journal->j_wbuf);
761         kfree(journal);
762         return NULL;
763 }
764
765 /**
766  *  journal_t * journal_init_inode () - creates a journal which maps to a inode.
767  *  @inode: An inode to create the journal in
768  *
769  * journal_init_inode creates a journal which maps an on-disk inode as
770  * the journal.  The inode must exist already, must support bmap() and
771  * must have all data blocks preallocated.
772  */
773 journal_t * journal_init_inode (struct inode *inode)
774 {
775         struct buffer_head *bh;
776         journal_t *journal = journal_init_common();
777         int err;
778         int n;
779         unsigned int blocknr;
780
781         if (!journal)
782                 return NULL;
783
784         journal->j_dev = journal->j_fs_dev = inode->i_sb->s_bdev;
785         journal->j_inode = inode;
786         jbd_debug(1,
787                   "journal %p: inode %s/%ld, size %Ld, bits %d, blksize %ld\n",
788                   journal, inode->i_sb->s_id, inode->i_ino,
789                   (long long) inode->i_size,
790                   inode->i_sb->s_blocksize_bits, inode->i_sb->s_blocksize);
791
792         journal->j_maxlen = inode->i_size >> inode->i_sb->s_blocksize_bits;
793         journal->j_blocksize = inode->i_sb->s_blocksize;
794
795         /* journal descriptor can store up to n blocks -bzzz */
796         n = journal->j_blocksize / sizeof(journal_block_tag_t);
797         journal->j_wbufsize = n;
798         journal->j_wbuf = kmalloc(n * sizeof(struct buffer_head*), GFP_KERNEL);
799         if (!journal->j_wbuf) {
800                 printk(KERN_ERR "%s: Cant allocate bhs for commit thread\n",
801                         __func__);
802                 goto out_err;
803         }
804
805         err = journal_bmap(journal, 0, &blocknr);
806         /* If that failed, give up */
807         if (err) {
808                 printk(KERN_ERR "%s: Cannnot locate journal superblock\n",
809                        __func__);
810                 goto out_err;
811         }
812
813         bh = __getblk(journal->j_dev, blocknr, journal->j_blocksize);
814         if (!bh) {
815                 printk(KERN_ERR
816                        "%s: Cannot get buffer for journal superblock\n",
817                        __func__);
818                 goto out_err;
819         }
820         journal->j_sb_buffer = bh;
821         journal->j_superblock = (journal_superblock_t *)bh->b_data;
822
823         return journal;
824 out_err:
825         kfree(journal->j_wbuf);
826         kfree(journal);
827         return NULL;
828 }
829
830 /*
831  * If the journal init or create aborts, we need to mark the journal
832  * superblock as being NULL to prevent the journal destroy from writing
833  * back a bogus superblock.
834  */
835 static void journal_fail_superblock (journal_t *journal)
836 {
837         struct buffer_head *bh = journal->j_sb_buffer;
838         brelse(bh);
839         journal->j_sb_buffer = NULL;
840 }
841
842 /*
843  * Given a journal_t structure, initialise the various fields for
844  * startup of a new journaling session.  We use this both when creating
845  * a journal, and after recovering an old journal to reset it for
846  * subsequent use.
847  */
848
849 static int journal_reset(journal_t *journal)
850 {
851         journal_superblock_t *sb = journal->j_superblock;
852         unsigned int first, last;
853
854         first = be32_to_cpu(sb->s_first);
855         last = be32_to_cpu(sb->s_maxlen);
856         if (first + JFS_MIN_JOURNAL_BLOCKS > last + 1) {
857                 printk(KERN_ERR "JBD: Journal too short (blocks %u-%u).\n",
858                        first, last);
859                 journal_fail_superblock(journal);
860                 return -EINVAL;
861         }
862
863         journal->j_first = first;
864         journal->j_last = last;
865
866         journal->j_head = first;
867         journal->j_tail = first;
868         journal->j_free = last - first;
869
870         journal->j_tail_sequence = journal->j_transaction_sequence;
871         journal->j_commit_sequence = journal->j_transaction_sequence - 1;
872         journal->j_commit_request = journal->j_commit_sequence;
873
874         journal->j_max_transaction_buffers = journal->j_maxlen / 4;
875
876         /* Add the dynamic fields and write it to disk. */
877         journal_update_superblock(journal, 1);
878         return journal_start_thread(journal);
879 }
880
881 /**
882  * int journal_create() - Initialise the new journal file
883  * @journal: Journal to create. This structure must have been initialised
884  *
885  * Given a journal_t structure which tells us which disk blocks we can
886  * use, create a new journal superblock and initialise all of the
887  * journal fields from scratch.
888  **/
889 int journal_create(journal_t *journal)
890 {
891         unsigned int blocknr;
892         struct buffer_head *bh;
893         journal_superblock_t *sb;
894         int i, err;
895
896         if (journal->j_maxlen < JFS_MIN_JOURNAL_BLOCKS) {
897                 printk (KERN_ERR "Journal length (%d blocks) too short.\n",
898                         journal->j_maxlen);
899                 journal_fail_superblock(journal);
900                 return -EINVAL;
901         }
902
903         if (journal->j_inode == NULL) {
904                 /*
905                  * We don't know what block to start at!
906                  */
907                 printk(KERN_EMERG
908                        "%s: creation of journal on external device!\n",
909                        __func__);
910                 BUG();
911         }
912
913         /* Zero out the entire journal on disk.  We cannot afford to
914            have any blocks on disk beginning with JFS_MAGIC_NUMBER. */
915         jbd_debug(1, "JBD: Zeroing out journal blocks...\n");
916         for (i = 0; i < journal->j_maxlen; i++) {
917                 err = journal_bmap(journal, i, &blocknr);
918                 if (err)
919                         return err;
920                 bh = __getblk(journal->j_dev, blocknr, journal->j_blocksize);
921                 lock_buffer(bh);
922                 memset (bh->b_data, 0, journal->j_blocksize);
923                 BUFFER_TRACE(bh, "marking dirty");
924                 mark_buffer_dirty(bh);
925                 BUFFER_TRACE(bh, "marking uptodate");
926                 set_buffer_uptodate(bh);
927                 unlock_buffer(bh);
928                 __brelse(bh);
929         }
930
931         sync_blockdev(journal->j_dev);
932         jbd_debug(1, "JBD: journal cleared.\n");
933
934         /* OK, fill in the initial static fields in the new superblock */
935         sb = journal->j_superblock;
936
937         sb->s_header.h_magic     = cpu_to_be32(JFS_MAGIC_NUMBER);
938         sb->s_header.h_blocktype = cpu_to_be32(JFS_SUPERBLOCK_V2);
939
940         sb->s_blocksize = cpu_to_be32(journal->j_blocksize);
941         sb->s_maxlen    = cpu_to_be32(journal->j_maxlen);
942         sb->s_first     = cpu_to_be32(1);
943
944         journal->j_transaction_sequence = 1;
945
946         journal->j_flags &= ~JFS_ABORT;
947         journal->j_format_version = 2;
948
949         return journal_reset(journal);
950 }
951
952 /**
953  * void journal_update_superblock() - Update journal sb on disk.
954  * @journal: The journal to update.
955  * @wait: Set to '0' if you don't want to wait for IO completion.
956  *
957  * Update a journal's dynamic superblock fields and write it to disk,
958  * optionally waiting for the IO to complete.
959  */
960 void journal_update_superblock(journal_t *journal, int wait)
961 {
962         journal_superblock_t *sb = journal->j_superblock;
963         struct buffer_head *bh = journal->j_sb_buffer;
964
965         /*
966          * As a special case, if the on-disk copy is already marked as needing
967          * no recovery (s_start == 0) and there are no outstanding transactions
968          * in the filesystem, then we can safely defer the superblock update
969          * until the next commit by setting JFS_FLUSHED.  This avoids
970          * attempting a write to a potential-readonly device.
971          */
972         if (sb->s_start == 0 && journal->j_tail_sequence ==
973                                 journal->j_transaction_sequence) {
974                 jbd_debug(1,"JBD: Skipping superblock update on recovered sb "
975                         "(start %u, seq %d, errno %d)\n",
976                         journal->j_tail, journal->j_tail_sequence,
977                         journal->j_errno);
978                 goto out;
979         }
980
981         spin_lock(&journal->j_state_lock);
982         jbd_debug(1,"JBD: updating superblock (start %u, seq %d, errno %d)\n",
983                   journal->j_tail, journal->j_tail_sequence, journal->j_errno);
984
985         sb->s_sequence = cpu_to_be32(journal->j_tail_sequence);
986         sb->s_start    = cpu_to_be32(journal->j_tail);
987         sb->s_errno    = cpu_to_be32(journal->j_errno);
988         spin_unlock(&journal->j_state_lock);
989
990         BUFFER_TRACE(bh, "marking dirty");
991         mark_buffer_dirty(bh);
992         if (wait)
993                 sync_dirty_buffer(bh);
994         else
995                 ll_rw_block(SWRITE, 1, &bh);
996
997 out:
998         /* If we have just flushed the log (by marking s_start==0), then
999          * any future commit will have to be careful to update the
1000          * superblock again to re-record the true start of the log. */
1001
1002         spin_lock(&journal->j_state_lock);
1003         if (sb->s_start)
1004                 journal->j_flags &= ~JFS_FLUSHED;
1005         else
1006                 journal->j_flags |= JFS_FLUSHED;
1007         spin_unlock(&journal->j_state_lock);
1008 }
1009
1010 /*
1011  * Read the superblock for a given journal, performing initial
1012  * validation of the format.
1013  */
1014
1015 static int journal_get_superblock(journal_t *journal)
1016 {
1017         struct buffer_head *bh;
1018         journal_superblock_t *sb;
1019         int err = -EIO;
1020
1021         bh = journal->j_sb_buffer;
1022
1023         J_ASSERT(bh != NULL);
1024         if (!buffer_uptodate(bh)) {
1025                 ll_rw_block(READ, 1, &bh);
1026                 wait_on_buffer(bh);
1027                 if (!buffer_uptodate(bh)) {
1028                         printk (KERN_ERR
1029                                 "JBD: IO error reading journal superblock\n");
1030                         goto out;
1031                 }
1032         }
1033
1034         sb = journal->j_superblock;
1035
1036         err = -EINVAL;
1037
1038         if (sb->s_header.h_magic != cpu_to_be32(JFS_MAGIC_NUMBER) ||
1039             sb->s_blocksize != cpu_to_be32(journal->j_blocksize)) {
1040                 printk(KERN_WARNING "JBD: no valid journal superblock found\n");
1041                 goto out;
1042         }
1043
1044         switch(be32_to_cpu(sb->s_header.h_blocktype)) {
1045         case JFS_SUPERBLOCK_V1:
1046                 journal->j_format_version = 1;
1047                 break;
1048         case JFS_SUPERBLOCK_V2:
1049                 journal->j_format_version = 2;
1050                 break;
1051         default:
1052                 printk(KERN_WARNING "JBD: unrecognised superblock format ID\n");
1053                 goto out;
1054         }
1055
1056         if (be32_to_cpu(sb->s_maxlen) < journal->j_maxlen)
1057                 journal->j_maxlen = be32_to_cpu(sb->s_maxlen);
1058         else if (be32_to_cpu(sb->s_maxlen) > journal->j_maxlen) {
1059                 printk (KERN_WARNING "JBD: journal file too short\n");
1060                 goto out;
1061         }
1062
1063         return 0;
1064
1065 out:
1066         journal_fail_superblock(journal);
1067         return err;
1068 }
1069
1070 /*
1071  * Load the on-disk journal superblock and read the key fields into the
1072  * journal_t.
1073  */
1074
1075 static int load_superblock(journal_t *journal)
1076 {
1077         int err;
1078         journal_superblock_t *sb;
1079
1080         err = journal_get_superblock(journal);
1081         if (err)
1082                 return err;
1083
1084         sb = journal->j_superblock;
1085
1086         journal->j_tail_sequence = be32_to_cpu(sb->s_sequence);
1087         journal->j_tail = be32_to_cpu(sb->s_start);
1088         journal->j_first = be32_to_cpu(sb->s_first);
1089         journal->j_last = be32_to_cpu(sb->s_maxlen);
1090         journal->j_errno = be32_to_cpu(sb->s_errno);
1091
1092         return 0;
1093 }
1094
1095
1096 /**
1097  * int journal_load() - Read journal from disk.
1098  * @journal: Journal to act on.
1099  *
1100  * Given a journal_t structure which tells us which disk blocks contain
1101  * a journal, read the journal from disk to initialise the in-memory
1102  * structures.
1103  */
1104 int journal_load(journal_t *journal)
1105 {
1106         int err;
1107         journal_superblock_t *sb;
1108
1109         err = load_superblock(journal);
1110         if (err)
1111                 return err;
1112
1113         sb = journal->j_superblock;
1114         /* If this is a V2 superblock, then we have to check the
1115          * features flags on it. */
1116
1117         if (journal->j_format_version >= 2) {
1118                 if ((sb->s_feature_ro_compat &
1119                      ~cpu_to_be32(JFS_KNOWN_ROCOMPAT_FEATURES)) ||
1120                     (sb->s_feature_incompat &
1121                      ~cpu_to_be32(JFS_KNOWN_INCOMPAT_FEATURES))) {
1122                         printk (KERN_WARNING
1123                                 "JBD: Unrecognised features on journal\n");
1124                         return -EINVAL;
1125                 }
1126         }
1127
1128         /* Let the recovery code check whether it needs to recover any
1129          * data from the journal. */
1130         if (journal_recover(journal))
1131                 goto recovery_error;
1132
1133         /* OK, we've finished with the dynamic journal bits:
1134          * reinitialise the dynamic contents of the superblock in memory
1135          * and reset them on disk. */
1136         if (journal_reset(journal))
1137                 goto recovery_error;
1138
1139         journal->j_flags &= ~JFS_ABORT;
1140         journal->j_flags |= JFS_LOADED;
1141         return 0;
1142
1143 recovery_error:
1144         printk (KERN_WARNING "JBD: recovery failed\n");
1145         return -EIO;
1146 }
1147
1148 /**
1149  * void journal_destroy() - Release a journal_t structure.
1150  * @journal: Journal to act on.
1151  *
1152  * Release a journal_t structure once it is no longer in use by the
1153  * journaled object.
1154  * Return <0 if we couldn't clean up the journal.
1155  */
1156 int journal_destroy(journal_t *journal)
1157 {
1158         int err = 0;
1159
1160         /* Wait for the commit thread to wake up and die. */
1161         journal_kill_thread(journal);
1162
1163         /* Force a final log commit */
1164         if (journal->j_running_transaction)
1165                 journal_commit_transaction(journal);
1166
1167         /* Force any old transactions to disk */
1168
1169         /* Totally anal locking here... */
1170         spin_lock(&journal->j_list_lock);
1171         while (journal->j_checkpoint_transactions != NULL) {
1172                 spin_unlock(&journal->j_list_lock);
1173                 log_do_checkpoint(journal);
1174                 spin_lock(&journal->j_list_lock);
1175         }
1176
1177         J_ASSERT(journal->j_running_transaction == NULL);
1178         J_ASSERT(journal->j_committing_transaction == NULL);
1179         J_ASSERT(journal->j_checkpoint_transactions == NULL);
1180         spin_unlock(&journal->j_list_lock);
1181
1182         if (journal->j_sb_buffer) {
1183                 if (!is_journal_aborted(journal)) {
1184                         /* We can now mark the journal as empty. */
1185                         journal->j_tail = 0;
1186                         journal->j_tail_sequence =
1187                                 ++journal->j_transaction_sequence;
1188                         journal_update_superblock(journal, 1);
1189                 } else {
1190                         err = -EIO;
1191                 }
1192                 brelse(journal->j_sb_buffer);
1193         }
1194
1195         if (journal->j_inode)
1196                 iput(journal->j_inode);
1197         if (journal->j_revoke)
1198                 journal_destroy_revoke(journal);
1199         kfree(journal->j_wbuf);
1200         kfree(journal);
1201
1202         return err;
1203 }
1204
1205
1206 /**
1207  *int journal_check_used_features () - Check if features specified are used.
1208  * @journal: Journal to check.
1209  * @compat: bitmask of compatible features
1210  * @ro: bitmask of features that force read-only mount
1211  * @incompat: bitmask of incompatible features
1212  *
1213  * Check whether the journal uses all of a given set of
1214  * features.  Return true (non-zero) if it does.
1215  **/
1216
1217 int journal_check_used_features (journal_t *journal, unsigned long compat,
1218                                  unsigned long ro, unsigned long incompat)
1219 {
1220         journal_superblock_t *sb;
1221
1222         if (!compat && !ro && !incompat)
1223                 return 1;
1224         if (journal->j_format_version == 1)
1225                 return 0;
1226
1227         sb = journal->j_superblock;
1228
1229         if (((be32_to_cpu(sb->s_feature_compat) & compat) == compat) &&
1230             ((be32_to_cpu(sb->s_feature_ro_compat) & ro) == ro) &&
1231             ((be32_to_cpu(sb->s_feature_incompat) & incompat) == incompat))
1232                 return 1;
1233
1234         return 0;
1235 }
1236
1237 /**
1238  * int journal_check_available_features() - Check feature set in journalling layer
1239  * @journal: Journal to check.
1240  * @compat: bitmask of compatible features
1241  * @ro: bitmask of features that force read-only mount
1242  * @incompat: bitmask of incompatible features
1243  *
1244  * Check whether the journaling code supports the use of
1245  * all of a given set of features on this journal.  Return true
1246  * (non-zero) if it can. */
1247
1248 int journal_check_available_features (journal_t *journal, unsigned long compat,
1249                                       unsigned long ro, unsigned long incompat)
1250 {
1251         journal_superblock_t *sb;
1252
1253         if (!compat && !ro && !incompat)
1254                 return 1;
1255
1256         sb = journal->j_superblock;
1257
1258         /* We can support any known requested features iff the
1259          * superblock is in version 2.  Otherwise we fail to support any
1260          * extended sb features. */
1261
1262         if (journal->j_format_version != 2)
1263                 return 0;
1264
1265         if ((compat   & JFS_KNOWN_COMPAT_FEATURES) == compat &&
1266             (ro       & JFS_KNOWN_ROCOMPAT_FEATURES) == ro &&
1267             (incompat & JFS_KNOWN_INCOMPAT_FEATURES) == incompat)
1268                 return 1;
1269
1270         return 0;
1271 }
1272
1273 /**
1274  * int journal_set_features () - Mark a given journal feature in the superblock
1275  * @journal: Journal to act on.
1276  * @compat: bitmask of compatible features
1277  * @ro: bitmask of features that force read-only mount
1278  * @incompat: bitmask of incompatible features
1279  *
1280  * Mark a given journal feature as present on the
1281  * superblock.  Returns true if the requested features could be set.
1282  *
1283  */
1284
1285 int journal_set_features (journal_t *journal, unsigned long compat,
1286                           unsigned long ro, unsigned long incompat)
1287 {
1288         journal_superblock_t *sb;
1289
1290         if (journal_check_used_features(journal, compat, ro, incompat))
1291                 return 1;
1292
1293         if (!journal_check_available_features(journal, compat, ro, incompat))
1294                 return 0;
1295
1296         jbd_debug(1, "Setting new features 0x%lx/0x%lx/0x%lx\n",
1297                   compat, ro, incompat);
1298
1299         sb = journal->j_superblock;
1300
1301         sb->s_feature_compat    |= cpu_to_be32(compat);
1302         sb->s_feature_ro_compat |= cpu_to_be32(ro);
1303         sb->s_feature_incompat  |= cpu_to_be32(incompat);
1304
1305         return 1;
1306 }
1307
1308
1309 /**
1310  * int journal_update_format () - Update on-disk journal structure.
1311  * @journal: Journal to act on.
1312  *
1313  * Given an initialised but unloaded journal struct, poke about in the
1314  * on-disk structure to update it to the most recent supported version.
1315  */
1316 int journal_update_format (journal_t *journal)
1317 {
1318         journal_superblock_t *sb;
1319         int err;
1320
1321         err = journal_get_superblock(journal);
1322         if (err)
1323                 return err;
1324
1325         sb = journal->j_superblock;
1326
1327         switch (be32_to_cpu(sb->s_header.h_blocktype)) {
1328         case JFS_SUPERBLOCK_V2:
1329                 return 0;
1330         case JFS_SUPERBLOCK_V1:
1331                 return journal_convert_superblock_v1(journal, sb);
1332         default:
1333                 break;
1334         }
1335         return -EINVAL;
1336 }
1337
1338 static int journal_convert_superblock_v1(journal_t *journal,
1339                                          journal_superblock_t *sb)
1340 {
1341         int offset, blocksize;
1342         struct buffer_head *bh;
1343
1344         printk(KERN_WARNING
1345                 "JBD: Converting superblock from version 1 to 2.\n");
1346
1347         /* Pre-initialise new fields to zero */
1348         offset = ((char *) &(sb->s_feature_compat)) - ((char *) sb);
1349         blocksize = be32_to_cpu(sb->s_blocksize);
1350         memset(&sb->s_feature_compat, 0, blocksize-offset);
1351
1352         sb->s_nr_users = cpu_to_be32(1);
1353         sb->s_header.h_blocktype = cpu_to_be32(JFS_SUPERBLOCK_V2);
1354         journal->j_format_version = 2;
1355
1356         bh = journal->j_sb_buffer;
1357         BUFFER_TRACE(bh, "marking dirty");
1358         mark_buffer_dirty(bh);
1359         sync_dirty_buffer(bh);
1360         return 0;
1361 }
1362
1363
1364 /**
1365  * int journal_flush () - Flush journal
1366  * @journal: Journal to act on.
1367  *
1368  * Flush all data for a given journal to disk and empty the journal.
1369  * Filesystems can use this when remounting readonly to ensure that
1370  * recovery does not need to happen on remount.
1371  */
1372
1373 int journal_flush(journal_t *journal)
1374 {
1375         int err = 0;
1376         transaction_t *transaction = NULL;
1377         unsigned int old_tail;
1378
1379         spin_lock(&journal->j_state_lock);
1380
1381         /* Force everything buffered to the log... */
1382         if (journal->j_running_transaction) {
1383                 transaction = journal->j_running_transaction;
1384                 __log_start_commit(journal, transaction->t_tid);
1385         } else if (journal->j_committing_transaction)
1386                 transaction = journal->j_committing_transaction;
1387
1388         /* Wait for the log commit to complete... */
1389         if (transaction) {
1390                 tid_t tid = transaction->t_tid;
1391
1392                 spin_unlock(&journal->j_state_lock);
1393                 log_wait_commit(journal, tid);
1394         } else {
1395                 spin_unlock(&journal->j_state_lock);
1396         }
1397
1398         /* ...and flush everything in the log out to disk. */
1399         spin_lock(&journal->j_list_lock);
1400         while (!err && journal->j_checkpoint_transactions != NULL) {
1401                 spin_unlock(&journal->j_list_lock);
1402                 mutex_lock(&journal->j_checkpoint_mutex);
1403                 err = log_do_checkpoint(journal);
1404                 mutex_unlock(&journal->j_checkpoint_mutex);
1405                 spin_lock(&journal->j_list_lock);
1406         }
1407         spin_unlock(&journal->j_list_lock);
1408
1409         if (is_journal_aborted(journal))
1410                 return -EIO;
1411
1412         cleanup_journal_tail(journal);
1413
1414         /* Finally, mark the journal as really needing no recovery.
1415          * This sets s_start==0 in the underlying superblock, which is
1416          * the magic code for a fully-recovered superblock.  Any future
1417          * commits of data to the journal will restore the current
1418          * s_start value. */
1419         spin_lock(&journal->j_state_lock);
1420         old_tail = journal->j_tail;
1421         journal->j_tail = 0;
1422         spin_unlock(&journal->j_state_lock);
1423         journal_update_superblock(journal, 1);
1424         spin_lock(&journal->j_state_lock);
1425         journal->j_tail = old_tail;
1426
1427         J_ASSERT(!journal->j_running_transaction);
1428         J_ASSERT(!journal->j_committing_transaction);
1429         J_ASSERT(!journal->j_checkpoint_transactions);
1430         J_ASSERT(journal->j_head == journal->j_tail);
1431         J_ASSERT(journal->j_tail_sequence == journal->j_transaction_sequence);
1432         spin_unlock(&journal->j_state_lock);
1433         return 0;
1434 }
1435
1436 /**
1437  * int journal_wipe() - Wipe journal contents
1438  * @journal: Journal to act on.
1439  * @write: flag (see below)
1440  *
1441  * Wipe out all of the contents of a journal, safely.  This will produce
1442  * a warning if the journal contains any valid recovery information.
1443  * Must be called between journal_init_*() and journal_load().
1444  *
1445  * If 'write' is non-zero, then we wipe out the journal on disk; otherwise
1446  * we merely suppress recovery.
1447  */
1448
1449 int journal_wipe(journal_t *journal, int write)
1450 {
1451         journal_superblock_t *sb;
1452         int err = 0;
1453
1454         J_ASSERT (!(journal->j_flags & JFS_LOADED));
1455
1456         err = load_superblock(journal);
1457         if (err)
1458                 return err;
1459
1460         sb = journal->j_superblock;
1461
1462         if (!journal->j_tail)
1463                 goto no_recovery;
1464
1465         printk (KERN_WARNING "JBD: %s recovery information on journal\n",
1466                 write ? "Clearing" : "Ignoring");
1467
1468         err = journal_skip_recovery(journal);
1469         if (write)
1470                 journal_update_superblock(journal, 1);
1471
1472  no_recovery:
1473         return err;
1474 }
1475
1476 /*
1477  * journal_dev_name: format a character string to describe on what
1478  * device this journal is present.
1479  */
1480
1481 static const char *journal_dev_name(journal_t *journal, char *buffer)
1482 {
1483         struct block_device *bdev;
1484
1485         if (journal->j_inode)
1486                 bdev = journal->j_inode->i_sb->s_bdev;
1487         else
1488                 bdev = journal->j_dev;
1489
1490         return bdevname(bdev, buffer);
1491 }
1492
1493 /*
1494  * Journal abort has very specific semantics, which we describe
1495  * for journal abort.
1496  *
1497  * Two internal function, which provide abort to te jbd layer
1498  * itself are here.
1499  */
1500
1501 /*
1502  * Quick version for internal journal use (doesn't lock the journal).
1503  * Aborts hard --- we mark the abort as occurred, but do _nothing_ else,
1504  * and don't attempt to make any other journal updates.
1505  */
1506 static void __journal_abort_hard(journal_t *journal)
1507 {
1508         transaction_t *transaction;
1509         char b[BDEVNAME_SIZE];
1510
1511         if (journal->j_flags & JFS_ABORT)
1512                 return;
1513
1514         printk(KERN_ERR "Aborting journal on device %s.\n",
1515                 journal_dev_name(journal, b));
1516
1517         spin_lock(&journal->j_state_lock);
1518         journal->j_flags |= JFS_ABORT;
1519         transaction = journal->j_running_transaction;
1520         if (transaction)
1521                 __log_start_commit(journal, transaction->t_tid);
1522         spin_unlock(&journal->j_state_lock);
1523 }
1524
1525 /* Soft abort: record the abort error status in the journal superblock,
1526  * but don't do any other IO. */
1527 static void __journal_abort_soft (journal_t *journal, int errno)
1528 {
1529         if (journal->j_flags & JFS_ABORT)
1530                 return;
1531
1532         if (!journal->j_errno)
1533                 journal->j_errno = errno;
1534
1535         __journal_abort_hard(journal);
1536
1537         if (errno)
1538                 journal_update_superblock(journal, 1);
1539 }
1540
1541 /**
1542  * void journal_abort () - Shutdown the journal immediately.
1543  * @journal: the journal to shutdown.
1544  * @errno:   an error number to record in the journal indicating
1545  *           the reason for the shutdown.
1546  *
1547  * Perform a complete, immediate shutdown of the ENTIRE
1548  * journal (not of a single transaction).  This operation cannot be
1549  * undone without closing and reopening the journal.
1550  *
1551  * The journal_abort function is intended to support higher level error
1552  * recovery mechanisms such as the ext2/ext3 remount-readonly error
1553  * mode.
1554  *
1555  * Journal abort has very specific semantics.  Any existing dirty,
1556  * unjournaled buffers in the main filesystem will still be written to
1557  * disk by bdflush, but the journaling mechanism will be suspended
1558  * immediately and no further transaction commits will be honoured.
1559  *
1560  * Any dirty, journaled buffers will be written back to disk without
1561  * hitting the journal.  Atomicity cannot be guaranteed on an aborted
1562  * filesystem, but we _do_ attempt to leave as much data as possible
1563  * behind for fsck to use for cleanup.
1564  *
1565  * Any attempt to get a new transaction handle on a journal which is in
1566  * ABORT state will just result in an -EROFS error return.  A
1567  * journal_stop on an existing handle will return -EIO if we have
1568  * entered abort state during the update.
1569  *
1570  * Recursive transactions are not disturbed by journal abort until the
1571  * final journal_stop, which will receive the -EIO error.
1572  *
1573  * Finally, the journal_abort call allows the caller to supply an errno
1574  * which will be recorded (if possible) in the journal superblock.  This
1575  * allows a client to record failure conditions in the middle of a
1576  * transaction without having to complete the transaction to record the
1577  * failure to disk.  ext3_error, for example, now uses this
1578  * functionality.
1579  *
1580  * Errors which originate from within the journaling layer will NOT
1581  * supply an errno; a null errno implies that absolutely no further
1582  * writes are done to the journal (unless there are any already in
1583  * progress).
1584  *
1585  */
1586
1587 void journal_abort(journal_t *journal, int errno)
1588 {
1589         __journal_abort_soft(journal, errno);
1590 }
1591
1592 /**
1593  * int journal_errno () - returns the journal's error state.
1594  * @journal: journal to examine.
1595  *
1596  * This is the errno numbet set with journal_abort(), the last
1597  * time the journal was mounted - if the journal was stopped
1598  * without calling abort this will be 0.
1599  *
1600  * If the journal has been aborted on this mount time -EROFS will
1601  * be returned.
1602  */
1603 int journal_errno(journal_t *journal)
1604 {
1605         int err;
1606
1607         spin_lock(&journal->j_state_lock);
1608         if (journal->j_flags & JFS_ABORT)
1609                 err = -EROFS;
1610         else
1611                 err = journal->j_errno;
1612         spin_unlock(&journal->j_state_lock);
1613         return err;
1614 }
1615
1616 /**
1617  * int journal_clear_err () - clears the journal's error state
1618  * @journal: journal to act on.
1619  *
1620  * An error must be cleared or Acked to take a FS out of readonly
1621  * mode.
1622  */
1623 int journal_clear_err(journal_t *journal)
1624 {
1625         int err = 0;
1626
1627         spin_lock(&journal->j_state_lock);
1628         if (journal->j_flags & JFS_ABORT)
1629                 err = -EROFS;
1630         else
1631                 journal->j_errno = 0;
1632         spin_unlock(&journal->j_state_lock);
1633         return err;
1634 }
1635
1636 /**
1637  * void journal_ack_err() - Ack journal err.
1638  * @journal: journal to act on.
1639  *
1640  * An error must be cleared or Acked to take a FS out of readonly
1641  * mode.
1642  */
1643 void journal_ack_err(journal_t *journal)
1644 {
1645         spin_lock(&journal->j_state_lock);
1646         if (journal->j_errno)
1647                 journal->j_flags |= JFS_ACK_ERR;
1648         spin_unlock(&journal->j_state_lock);
1649 }
1650
1651 int journal_blocks_per_page(struct inode *inode)
1652 {
1653         return 1 << (PAGE_CACHE_SHIFT - inode->i_sb->s_blocksize_bits);
1654 }
1655
1656 /*
1657  * Journal_head storage management
1658  */
1659 static struct kmem_cache *journal_head_cache;
1660 #ifdef CONFIG_JBD_DEBUG
1661 static atomic_t nr_journal_heads = ATOMIC_INIT(0);
1662 #endif
1663
1664 static int journal_init_journal_head_cache(void)
1665 {
1666         int retval;
1667
1668         J_ASSERT(journal_head_cache == NULL);
1669         journal_head_cache = kmem_cache_create("journal_head",
1670                                 sizeof(struct journal_head),
1671                                 0,              /* offset */
1672                                 SLAB_TEMPORARY, /* flags */
1673                                 NULL);          /* ctor */
1674         retval = 0;
1675         if (!journal_head_cache) {
1676                 retval = -ENOMEM;
1677                 printk(KERN_EMERG "JBD: no memory for journal_head cache\n");
1678         }
1679         return retval;
1680 }
1681
1682 static void journal_destroy_journal_head_cache(void)
1683 {
1684         if (journal_head_cache) {
1685                 kmem_cache_destroy(journal_head_cache);
1686                 journal_head_cache = NULL;
1687         }
1688 }
1689
1690 /*
1691  * journal_head splicing and dicing
1692  */
1693 static struct journal_head *journal_alloc_journal_head(void)
1694 {
1695         struct journal_head *ret;
1696         static unsigned long last_warning;
1697
1698 #ifdef CONFIG_JBD_DEBUG
1699         atomic_inc(&nr_journal_heads);
1700 #endif
1701         ret = kmem_cache_alloc(journal_head_cache, GFP_NOFS);
1702         if (ret == NULL) {
1703                 jbd_debug(1, "out of memory for journal_head\n");
1704                 if (time_after(jiffies, last_warning + 5*HZ)) {
1705                         printk(KERN_NOTICE "ENOMEM in %s, retrying.\n",
1706                                __func__);
1707                         last_warning = jiffies;
1708                 }
1709                 while (ret == NULL) {
1710                         yield();
1711                         ret = kmem_cache_alloc(journal_head_cache, GFP_NOFS);
1712                 }
1713         }
1714         return ret;
1715 }
1716
1717 static void journal_free_journal_head(struct journal_head *jh)
1718 {
1719 #ifdef CONFIG_JBD_DEBUG
1720         atomic_dec(&nr_journal_heads);
1721         memset(jh, JBD_POISON_FREE, sizeof(*jh));
1722 #endif
1723         kmem_cache_free(journal_head_cache, jh);
1724 }
1725
1726 /*
1727  * A journal_head is attached to a buffer_head whenever JBD has an
1728  * interest in the buffer.
1729  *
1730  * Whenever a buffer has an attached journal_head, its ->b_state:BH_JBD bit
1731  * is set.  This bit is tested in core kernel code where we need to take
1732  * JBD-specific actions.  Testing the zeroness of ->b_private is not reliable
1733  * there.
1734  *
1735  * When a buffer has its BH_JBD bit set, its ->b_count is elevated by one.
1736  *
1737  * When a buffer has its BH_JBD bit set it is immune from being released by
1738  * core kernel code, mainly via ->b_count.
1739  *
1740  * A journal_head may be detached from its buffer_head when the journal_head's
1741  * b_transaction, b_cp_transaction and b_next_transaction pointers are NULL.
1742  * Various places in JBD call journal_remove_journal_head() to indicate that the
1743  * journal_head can be dropped if needed.
1744  *
1745  * Various places in the kernel want to attach a journal_head to a buffer_head
1746  * _before_ attaching the journal_head to a transaction.  To protect the
1747  * journal_head in this situation, journal_add_journal_head elevates the
1748  * journal_head's b_jcount refcount by one.  The caller must call
1749  * journal_put_journal_head() to undo this.
1750  *
1751  * So the typical usage would be:
1752  *
1753  *      (Attach a journal_head if needed.  Increments b_jcount)
1754  *      struct journal_head *jh = journal_add_journal_head(bh);
1755  *      ...
1756  *      jh->b_transaction = xxx;
1757  *      journal_put_journal_head(jh);
1758  *
1759  * Now, the journal_head's b_jcount is zero, but it is safe from being released
1760  * because it has a non-zero b_transaction.
1761  */
1762
1763 /*
1764  * Give a buffer_head a journal_head.
1765  *
1766  * Doesn't need the journal lock.
1767  * May sleep.
1768  */
1769 struct journal_head *journal_add_journal_head(struct buffer_head *bh)
1770 {
1771         struct journal_head *jh;
1772         struct journal_head *new_jh = NULL;
1773
1774 repeat:
1775         if (!buffer_jbd(bh)) {
1776                 new_jh = journal_alloc_journal_head();
1777                 memset(new_jh, 0, sizeof(*new_jh));
1778         }
1779
1780         jbd_lock_bh_journal_head(bh);
1781         if (buffer_jbd(bh)) {
1782                 jh = bh2jh(bh);
1783         } else {
1784                 J_ASSERT_BH(bh,
1785                         (atomic_read(&bh->b_count) > 0) ||
1786                         (bh->b_page && bh->b_page->mapping));
1787
1788                 if (!new_jh) {
1789                         jbd_unlock_bh_journal_head(bh);
1790                         goto repeat;
1791                 }
1792
1793                 jh = new_jh;
1794                 new_jh = NULL;          /* We consumed it */
1795                 set_buffer_jbd(bh);
1796                 bh->b_private = jh;
1797                 jh->b_bh = bh;
1798                 get_bh(bh);
1799                 BUFFER_TRACE(bh, "added journal_head");
1800         }
1801         jh->b_jcount++;
1802         jbd_unlock_bh_journal_head(bh);
1803         if (new_jh)
1804                 journal_free_journal_head(new_jh);
1805         return bh->b_private;
1806 }
1807
1808 /*
1809  * Grab a ref against this buffer_head's journal_head.  If it ended up not
1810  * having a journal_head, return NULL
1811  */
1812 struct journal_head *journal_grab_journal_head(struct buffer_head *bh)
1813 {
1814         struct journal_head *jh = NULL;
1815
1816         jbd_lock_bh_journal_head(bh);
1817         if (buffer_jbd(bh)) {
1818                 jh = bh2jh(bh);
1819                 jh->b_jcount++;
1820         }
1821         jbd_unlock_bh_journal_head(bh);
1822         return jh;
1823 }
1824
1825 static void __journal_remove_journal_head(struct buffer_head *bh)
1826 {
1827         struct journal_head *jh = bh2jh(bh);
1828
1829         J_ASSERT_JH(jh, jh->b_jcount >= 0);
1830
1831         get_bh(bh);
1832         if (jh->b_jcount == 0) {
1833                 if (jh->b_transaction == NULL &&
1834                                 jh->b_next_transaction == NULL &&
1835                                 jh->b_cp_transaction == NULL) {
1836                         J_ASSERT_JH(jh, jh->b_jlist == BJ_None);
1837                         J_ASSERT_BH(bh, buffer_jbd(bh));
1838                         J_ASSERT_BH(bh, jh2bh(jh) == bh);
1839                         BUFFER_TRACE(bh, "remove journal_head");
1840                         if (jh->b_frozen_data) {
1841                                 printk(KERN_WARNING "%s: freeing "
1842                                                 "b_frozen_data\n",
1843                                                 __func__);
1844                                 jbd_free(jh->b_frozen_data, bh->b_size);
1845                         }
1846                         if (jh->b_committed_data) {
1847                                 printk(KERN_WARNING "%s: freeing "
1848                                                 "b_committed_data\n",
1849                                                 __func__);
1850                                 jbd_free(jh->b_committed_data, bh->b_size);
1851                         }
1852                         bh->b_private = NULL;
1853                         jh->b_bh = NULL;        /* debug, really */
1854                         clear_buffer_jbd(bh);
1855                         __brelse(bh);
1856                         journal_free_journal_head(jh);
1857                 } else {
1858                         BUFFER_TRACE(bh, "journal_head was locked");
1859                 }
1860         }
1861 }
1862
1863 /*
1864  * journal_remove_journal_head(): if the buffer isn't attached to a transaction
1865  * and has a zero b_jcount then remove and release its journal_head.   If we did
1866  * see that the buffer is not used by any transaction we also "logically"
1867  * decrement ->b_count.
1868  *
1869  * We in fact take an additional increment on ->b_count as a convenience,
1870  * because the caller usually wants to do additional things with the bh
1871  * after calling here.
1872  * The caller of journal_remove_journal_head() *must* run __brelse(bh) at some
1873  * time.  Once the caller has run __brelse(), the buffer is eligible for
1874  * reaping by try_to_free_buffers().
1875  */
1876 void journal_remove_journal_head(struct buffer_head *bh)
1877 {
1878         jbd_lock_bh_journal_head(bh);
1879         __journal_remove_journal_head(bh);
1880         jbd_unlock_bh_journal_head(bh);
1881 }
1882
1883 /*
1884  * Drop a reference on the passed journal_head.  If it fell to zero then try to
1885  * release the journal_head from the buffer_head.
1886  */
1887 void journal_put_journal_head(struct journal_head *jh)
1888 {
1889         struct buffer_head *bh = jh2bh(jh);
1890
1891         jbd_lock_bh_journal_head(bh);
1892         J_ASSERT_JH(jh, jh->b_jcount > 0);
1893         --jh->b_jcount;
1894         if (!jh->b_jcount && !jh->b_transaction) {
1895                 __journal_remove_journal_head(bh);
1896                 __brelse(bh);
1897         }
1898         jbd_unlock_bh_journal_head(bh);
1899 }
1900
1901 /*
1902  * debugfs tunables
1903  */
1904 #ifdef CONFIG_JBD_DEBUG
1905
1906 u8 journal_enable_debug __read_mostly;
1907 EXPORT_SYMBOL(journal_enable_debug);
1908
1909 static struct dentry *jbd_debugfs_dir;
1910 static struct dentry *jbd_debug;
1911
1912 static void __init jbd_create_debugfs_entry(void)
1913 {
1914         jbd_debugfs_dir = debugfs_create_dir("jbd", NULL);
1915         if (jbd_debugfs_dir)
1916                 jbd_debug = debugfs_create_u8("jbd-debug", S_IRUGO | S_IWUSR,
1917                                                jbd_debugfs_dir,
1918                                                &journal_enable_debug);
1919 }
1920
1921 static void __exit jbd_remove_debugfs_entry(void)
1922 {
1923         debugfs_remove(jbd_debug);
1924         debugfs_remove(jbd_debugfs_dir);
1925 }
1926
1927 #else
1928
1929 static inline void jbd_create_debugfs_entry(void)
1930 {
1931 }
1932
1933 static inline void jbd_remove_debugfs_entry(void)
1934 {
1935 }
1936
1937 #endif
1938
1939 struct kmem_cache *jbd_handle_cache;
1940
1941 static int __init journal_init_handle_cache(void)
1942 {
1943         jbd_handle_cache = kmem_cache_create("journal_handle",
1944                                 sizeof(handle_t),
1945                                 0,              /* offset */
1946                                 SLAB_TEMPORARY, /* flags */
1947                                 NULL);          /* ctor */
1948         if (jbd_handle_cache == NULL) {
1949                 printk(KERN_EMERG "JBD: failed to create handle cache\n");
1950                 return -ENOMEM;
1951         }
1952         return 0;
1953 }
1954
1955 static void journal_destroy_handle_cache(void)
1956 {
1957         if (jbd_handle_cache)
1958                 kmem_cache_destroy(jbd_handle_cache);
1959 }
1960
1961 /*
1962  * Module startup and shutdown
1963  */
1964
1965 static int __init journal_init_caches(void)
1966 {
1967         int ret;
1968
1969         ret = journal_init_revoke_caches();
1970         if (ret == 0)
1971                 ret = journal_init_journal_head_cache();
1972         if (ret == 0)
1973                 ret = journal_init_handle_cache();
1974         return ret;
1975 }
1976
1977 static void journal_destroy_caches(void)
1978 {
1979         journal_destroy_revoke_caches();
1980         journal_destroy_journal_head_cache();
1981         journal_destroy_handle_cache();
1982 }
1983
1984 static int __init journal_init(void)
1985 {
1986         int ret;
1987
1988         BUILD_BUG_ON(sizeof(struct journal_superblock_s) != 1024);
1989
1990         ret = journal_init_caches();
1991         if (ret != 0)
1992                 journal_destroy_caches();
1993         jbd_create_debugfs_entry();
1994         return ret;
1995 }
1996
1997 static void __exit journal_exit(void)
1998 {
1999 #ifdef CONFIG_JBD_DEBUG
2000         int n = atomic_read(&nr_journal_heads);
2001         if (n)
2002                 printk(KERN_EMERG "JBD: leaked %d journal_heads!\n", n);
2003 #endif
2004         jbd_remove_debugfs_entry();
2005         journal_destroy_caches();
2006 }
2007
2008 MODULE_LICENSE("GPL");
2009 module_init(journal_init);
2010 module_exit(journal_exit);
2011