[PATCH] fuse: miscellaneous cleanup
[linux-2.6] / fs / jbd / journal.c
1 /*
2  * linux/fs/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/smp_lock.h>
32 #include <linux/init.h>
33 #include <linux/mm.h>
34 #include <linux/suspend.h>
35 #include <linux/pagemap.h>
36 #include <asm/uaccess.h>
37 #include <asm/page.h>
38 #include <linux/proc_fs.h>
39
40 EXPORT_SYMBOL(journal_start);
41 EXPORT_SYMBOL(journal_restart);
42 EXPORT_SYMBOL(journal_extend);
43 EXPORT_SYMBOL(journal_stop);
44 EXPORT_SYMBOL(journal_lock_updates);
45 EXPORT_SYMBOL(journal_unlock_updates);
46 EXPORT_SYMBOL(journal_get_write_access);
47 EXPORT_SYMBOL(journal_get_create_access);
48 EXPORT_SYMBOL(journal_get_undo_access);
49 EXPORT_SYMBOL(journal_dirty_data);
50 EXPORT_SYMBOL(journal_dirty_metadata);
51 EXPORT_SYMBOL(journal_release_buffer);
52 EXPORT_SYMBOL(journal_forget);
53 #if 0
54 EXPORT_SYMBOL(journal_sync_buffer);
55 #endif
56 EXPORT_SYMBOL(journal_flush);
57 EXPORT_SYMBOL(journal_revoke);
58
59 EXPORT_SYMBOL(journal_init_dev);
60 EXPORT_SYMBOL(journal_init_inode);
61 EXPORT_SYMBOL(journal_update_format);
62 EXPORT_SYMBOL(journal_check_used_features);
63 EXPORT_SYMBOL(journal_check_available_features);
64 EXPORT_SYMBOL(journal_set_features);
65 EXPORT_SYMBOL(journal_create);
66 EXPORT_SYMBOL(journal_load);
67 EXPORT_SYMBOL(journal_destroy);
68 EXPORT_SYMBOL(journal_update_superblock);
69 EXPORT_SYMBOL(journal_abort);
70 EXPORT_SYMBOL(journal_errno);
71 EXPORT_SYMBOL(journal_ack_err);
72 EXPORT_SYMBOL(journal_clear_err);
73 EXPORT_SYMBOL(log_wait_commit);
74 EXPORT_SYMBOL(journal_start_commit);
75 EXPORT_SYMBOL(journal_force_commit_nested);
76 EXPORT_SYMBOL(journal_wipe);
77 EXPORT_SYMBOL(journal_blocks_per_page);
78 EXPORT_SYMBOL(journal_invalidatepage);
79 EXPORT_SYMBOL(journal_try_to_free_buffers);
80 EXPORT_SYMBOL(journal_force_commit);
81
82 static int journal_convert_superblock_v1(journal_t *, journal_superblock_t *);
83 static void __journal_abort_soft (journal_t *journal, int errno);
84
85 /*
86  * Helper function used to manage commit timeouts
87  */
88
89 static void commit_timeout(unsigned long __data)
90 {
91         struct task_struct * p = (struct task_struct *) __data;
92
93         wake_up_process(p);
94 }
95
96 /*
97  * kjournald: The main thread function used to manage a logging device
98  * journal.
99  *
100  * This kernel thread is responsible for two things:
101  *
102  * 1) COMMIT:  Every so often we need to commit the current state of the
103  *    filesystem to disk.  The journal thread is responsible for writing
104  *    all of the metadata buffers to disk.
105  *
106  * 2) CHECKPOINT: We cannot reuse a used section of the log file until all
107  *    of the data in that part of the log has been rewritten elsewhere on
108  *    the disk.  Flushing these old buffers to reclaim space in the log is
109  *    known as checkpointing, and this thread is responsible for that job.
110  */
111
112 static int kjournald(void *arg)
113 {
114         journal_t *journal = (journal_t *) arg;
115         transaction_t *transaction;
116         struct timer_list timer;
117
118         daemonize("kjournald");
119
120         /* Set up an interval timer which can be used to trigger a
121            commit wakeup after the commit interval expires */
122         init_timer(&timer);
123         timer.data = (unsigned long) current;
124         timer.function = commit_timeout;
125         journal->j_commit_timer = &timer;
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 void journal_start_thread(journal_t *journal)
214 {
215         kernel_thread(kjournald, journal, CLONE_VM|CLONE_FS|CLONE_FILES);
216         wait_event(journal->j_wait_done_commit, journal->j_task != 0);
217 }
218
219 static void journal_kill_thread(journal_t *journal)
220 {
221         spin_lock(&journal->j_state_lock);
222         journal->j_flags |= JFS_UNMOUNT;
223
224         while (journal->j_task) {
225                 wake_up(&journal->j_wait_commit);
226                 spin_unlock(&journal->j_state_lock);
227                 wait_event(journal->j_wait_done_commit, journal->j_task == 0);
228                 spin_lock(&journal->j_state_lock);
229         }
230         spin_unlock(&journal->j_state_lock);
231 }
232
233 /*
234  * journal_write_metadata_buffer: write a metadata buffer to the journal.
235  *
236  * Writes a metadata buffer to a given disk block.  The actual IO is not
237  * performed but a new buffer_head is constructed which labels the data
238  * to be written with the correct destination disk block.
239  *
240  * Any magic-number escaping which needs to be done will cause a
241  * copy-out here.  If the buffer happens to start with the
242  * JFS_MAGIC_NUMBER, then we can't write it to the log directly: the
243  * magic number is only written to the log for descripter blocks.  In
244  * this case, we copy the data and replace the first word with 0, and we
245  * return a result code which indicates that this buffer needs to be
246  * marked as an escaped buffer in the corresponding log descriptor
247  * block.  The missing word can then be restored when the block is read
248  * during recovery.
249  *
250  * If the source buffer has already been modified by a new transaction
251  * since we took the last commit snapshot, we use the frozen copy of
252  * that data for IO.  If we end up using the existing buffer_head's data
253  * for the write, then we *have* to lock the buffer to prevent anyone
254  * else from using and possibly modifying it while the IO is in
255  * progress.
256  *
257  * The function returns a pointer to the buffer_heads to be used for IO.
258  *
259  * We assume that the journal has already been locked in this function.
260  *
261  * Return value:
262  *  <0: Error
263  * >=0: Finished OK
264  *
265  * On success:
266  * Bit 0 set == escape performed on the data
267  * Bit 1 set == buffer copy-out performed (kfree the data after IO)
268  */
269
270 int journal_write_metadata_buffer(transaction_t *transaction,
271                                   struct journal_head  *jh_in,
272                                   struct journal_head **jh_out,
273                                   int blocknr)
274 {
275         int need_copy_out = 0;
276         int done_copy_out = 0;
277         int do_escape = 0;
278         char *mapped_data;
279         struct buffer_head *new_bh;
280         struct journal_head *new_jh;
281         struct page *new_page;
282         unsigned int new_offset;
283         struct buffer_head *bh_in = jh2bh(jh_in);
284
285         /*
286          * The buffer really shouldn't be locked: only the current committing
287          * transaction is allowed to write it, so nobody else is allowed
288          * to do any IO.
289          *
290          * akpm: except if we're journalling data, and write() output is
291          * also part of a shared mapping, and another thread has
292          * decided to launch a writepage() against this buffer.
293          */
294         J_ASSERT_BH(bh_in, buffer_jbddirty(bh_in));
295
296         new_bh = alloc_buffer_head(GFP_NOFS|__GFP_NOFAIL);
297
298         /*
299          * If a new transaction has already done a buffer copy-out, then
300          * we use that version of the data for the commit.
301          */
302         jbd_lock_bh_state(bh_in);
303 repeat:
304         if (jh_in->b_frozen_data) {
305                 done_copy_out = 1;
306                 new_page = virt_to_page(jh_in->b_frozen_data);
307                 new_offset = offset_in_page(jh_in->b_frozen_data);
308         } else {
309                 new_page = jh2bh(jh_in)->b_page;
310                 new_offset = offset_in_page(jh2bh(jh_in)->b_data);
311         }
312
313         mapped_data = kmap_atomic(new_page, KM_USER0);
314         /*
315          * Check for escaping
316          */
317         if (*((__be32 *)(mapped_data + new_offset)) ==
318                                 cpu_to_be32(JFS_MAGIC_NUMBER)) {
319                 need_copy_out = 1;
320                 do_escape = 1;
321         }
322         kunmap_atomic(mapped_data, KM_USER0);
323
324         /*
325          * Do we need to do a data copy?
326          */
327         if (need_copy_out && !done_copy_out) {
328                 char *tmp;
329
330                 jbd_unlock_bh_state(bh_in);
331                 tmp = jbd_rep_kmalloc(bh_in->b_size, GFP_NOFS);
332                 jbd_lock_bh_state(bh_in);
333                 if (jh_in->b_frozen_data) {
334                         kfree(tmp);
335                         goto repeat;
336                 }
337
338                 jh_in->b_frozen_data = tmp;
339                 mapped_data = kmap_atomic(new_page, KM_USER0);
340                 memcpy(tmp, mapped_data + new_offset, jh2bh(jh_in)->b_size);
341                 kunmap_atomic(mapped_data, KM_USER0);
342
343                 new_page = virt_to_page(tmp);
344                 new_offset = offset_in_page(tmp);
345                 done_copy_out = 1;
346         }
347
348         /*
349          * Did we need to do an escaping?  Now we've done all the
350          * copying, we can finally do so.
351          */
352         if (do_escape) {
353                 mapped_data = kmap_atomic(new_page, KM_USER0);
354                 *((unsigned int *)(mapped_data + new_offset)) = 0;
355                 kunmap_atomic(mapped_data, KM_USER0);
356         }
357
358         /* keep subsequent assertions sane */
359         new_bh->b_state = 0;
360         init_buffer(new_bh, NULL, NULL);
361         atomic_set(&new_bh->b_count, 1);
362         jbd_unlock_bh_state(bh_in);
363
364         new_jh = journal_add_journal_head(new_bh);      /* This sleeps */
365
366         set_bh_page(new_bh, new_page, new_offset);
367         new_jh->b_transaction = NULL;
368         new_bh->b_size = jh2bh(jh_in)->b_size;
369         new_bh->b_bdev = transaction->t_journal->j_dev;
370         new_bh->b_blocknr = blocknr;
371         set_buffer_mapped(new_bh);
372         set_buffer_dirty(new_bh);
373
374         *jh_out = new_jh;
375
376         /*
377          * The to-be-written buffer needs to get moved to the io queue,
378          * and the original buffer whose contents we are shadowing or
379          * copying is moved to the transaction's shadow queue.
380          */
381         JBUFFER_TRACE(jh_in, "file as BJ_Shadow");
382         journal_file_buffer(jh_in, transaction, BJ_Shadow);
383         JBUFFER_TRACE(new_jh, "file as BJ_IO");
384         journal_file_buffer(new_jh, transaction, BJ_IO);
385
386         return do_escape | (done_copy_out << 1);
387 }
388
389 /*
390  * Allocation code for the journal file.  Manage the space left in the
391  * journal, so that we can begin checkpointing when appropriate.
392  */
393
394 /*
395  * __log_space_left: Return the number of free blocks left in the journal.
396  *
397  * Called with the journal already locked.
398  *
399  * Called under j_state_lock
400  */
401
402 int __log_space_left(journal_t *journal)
403 {
404         int left = journal->j_free;
405
406         assert_spin_locked(&journal->j_state_lock);
407
408         /*
409          * Be pessimistic here about the number of those free blocks which
410          * might be required for log descriptor control blocks.
411          */
412
413 #define MIN_LOG_RESERVED_BLOCKS 32 /* Allow for rounding errors */
414
415         left -= MIN_LOG_RESERVED_BLOCKS;
416
417         if (left <= 0)
418                 return 0;
419         left -= (left >> 3);
420         return left;
421 }
422
423 /*
424  * Called under j_state_lock.  Returns true if a transaction was started.
425  */
426 int __log_start_commit(journal_t *journal, tid_t target)
427 {
428         /*
429          * Are we already doing a recent enough commit?
430          */
431         if (!tid_geq(journal->j_commit_request, target)) {
432                 /*
433                  * We want a new commit: OK, mark the request and wakup the
434                  * commit thread.  We do _not_ do the commit ourselves.
435                  */
436
437                 journal->j_commit_request = target;
438                 jbd_debug(1, "JBD: requesting commit %d/%d\n",
439                           journal->j_commit_request,
440                           journal->j_commit_sequence);
441                 wake_up(&journal->j_wait_commit);
442                 return 1;
443         }
444         return 0;
445 }
446
447 int log_start_commit(journal_t *journal, tid_t tid)
448 {
449         int ret;
450
451         spin_lock(&journal->j_state_lock);
452         ret = __log_start_commit(journal, tid);
453         spin_unlock(&journal->j_state_lock);
454         return ret;
455 }
456
457 /*
458  * Force and wait upon a commit if the calling process is not within
459  * transaction.  This is used for forcing out undo-protected data which contains
460  * bitmaps, when the fs is running out of space.
461  *
462  * We can only force the running transaction if we don't have an active handle;
463  * otherwise, we will deadlock.
464  *
465  * Returns true if a transaction was started.
466  */
467 int journal_force_commit_nested(journal_t *journal)
468 {
469         transaction_t *transaction = NULL;
470         tid_t tid;
471
472         spin_lock(&journal->j_state_lock);
473         if (journal->j_running_transaction && !current->journal_info) {
474                 transaction = journal->j_running_transaction;
475                 __log_start_commit(journal, transaction->t_tid);
476         } else if (journal->j_committing_transaction)
477                 transaction = journal->j_committing_transaction;
478
479         if (!transaction) {
480                 spin_unlock(&journal->j_state_lock);
481                 return 0;       /* Nothing to retry */
482         }
483
484         tid = transaction->t_tid;
485         spin_unlock(&journal->j_state_lock);
486         log_wait_commit(journal, tid);
487         return 1;
488 }
489
490 /*
491  * Start a commit of the current running transaction (if any).  Returns true
492  * if a transaction was started, and fills its tid in at *ptid
493  */
494 int journal_start_commit(journal_t *journal, tid_t *ptid)
495 {
496         int ret = 0;
497
498         spin_lock(&journal->j_state_lock);
499         if (journal->j_running_transaction) {
500                 tid_t tid = journal->j_running_transaction->t_tid;
501
502                 ret = __log_start_commit(journal, tid);
503                 if (ret && ptid)
504                         *ptid = tid;
505         } else if (journal->j_committing_transaction && ptid) {
506                 /*
507                  * If ext3_write_super() recently started a commit, then we
508                  * have to wait for completion of that transaction
509                  */
510                 *ptid = journal->j_committing_transaction->t_tid;
511                 ret = 1;
512         }
513         spin_unlock(&journal->j_state_lock);
514         return ret;
515 }
516
517 /*
518  * Wait for a specified commit to complete.
519  * The caller may not hold the journal lock.
520  */
521 int log_wait_commit(journal_t *journal, tid_t tid)
522 {
523         int err = 0;
524
525 #ifdef CONFIG_JBD_DEBUG
526         spin_lock(&journal->j_state_lock);
527         if (!tid_geq(journal->j_commit_request, tid)) {
528                 printk(KERN_EMERG
529                        "%s: error: j_commit_request=%d, tid=%d\n",
530                        __FUNCTION__, journal->j_commit_request, tid);
531         }
532         spin_unlock(&journal->j_state_lock);
533 #endif
534         spin_lock(&journal->j_state_lock);
535         while (tid_gt(tid, journal->j_commit_sequence)) {
536                 jbd_debug(1, "JBD: want %d, j_commit_sequence=%d\n",
537                                   tid, journal->j_commit_sequence);
538                 wake_up(&journal->j_wait_commit);
539                 spin_unlock(&journal->j_state_lock);
540                 wait_event(journal->j_wait_done_commit,
541                                 !tid_gt(tid, journal->j_commit_sequence));
542                 spin_lock(&journal->j_state_lock);
543         }
544         spin_unlock(&journal->j_state_lock);
545
546         if (unlikely(is_journal_aborted(journal))) {
547                 printk(KERN_EMERG "journal commit I/O error\n");
548                 err = -EIO;
549         }
550         return err;
551 }
552
553 /*
554  * Log buffer allocation routines:
555  */
556
557 int journal_next_log_block(journal_t *journal, unsigned long *retp)
558 {
559         unsigned long blocknr;
560
561         spin_lock(&journal->j_state_lock);
562         J_ASSERT(journal->j_free > 1);
563
564         blocknr = journal->j_head;
565         journal->j_head++;
566         journal->j_free--;
567         if (journal->j_head == journal->j_last)
568                 journal->j_head = journal->j_first;
569         spin_unlock(&journal->j_state_lock);
570         return journal_bmap(journal, blocknr, retp);
571 }
572
573 /*
574  * Conversion of logical to physical block numbers for the journal
575  *
576  * On external journals the journal blocks are identity-mapped, so
577  * this is a no-op.  If needed, we can use j_blk_offset - everything is
578  * ready.
579  */
580 int journal_bmap(journal_t *journal, unsigned long blocknr, 
581                  unsigned long *retp)
582 {
583         int err = 0;
584         unsigned long ret;
585
586         if (journal->j_inode) {
587                 ret = bmap(journal->j_inode, blocknr);
588                 if (ret)
589                         *retp = ret;
590                 else {
591                         char b[BDEVNAME_SIZE];
592
593                         printk(KERN_ALERT "%s: journal block not found "
594                                         "at offset %lu on %s\n",
595                                 __FUNCTION__,
596                                 blocknr,
597                                 bdevname(journal->j_dev, b));
598                         err = -EIO;
599                         __journal_abort_soft(journal, err);
600                 }
601         } else {
602                 *retp = blocknr; /* +journal->j_blk_offset */
603         }
604         return err;
605 }
606
607 /*
608  * We play buffer_head aliasing tricks to write data/metadata blocks to
609  * the journal without copying their contents, but for journal
610  * descriptor blocks we do need to generate bona fide buffers.
611  *
612  * After the caller of journal_get_descriptor_buffer() has finished modifying
613  * the buffer's contents they really should run flush_dcache_page(bh->b_page).
614  * But we don't bother doing that, so there will be coherency problems with
615  * mmaps of blockdevs which hold live JBD-controlled filesystems.
616  */
617 struct journal_head *journal_get_descriptor_buffer(journal_t *journal)
618 {
619         struct buffer_head *bh;
620         unsigned long blocknr;
621         int err;
622
623         err = journal_next_log_block(journal, &blocknr);
624
625         if (err)
626                 return NULL;
627
628         bh = __getblk(journal->j_dev, blocknr, journal->j_blocksize);
629         lock_buffer(bh);
630         memset(bh->b_data, 0, journal->j_blocksize);
631         set_buffer_uptodate(bh);
632         unlock_buffer(bh);
633         BUFFER_TRACE(bh, "return this buffer");
634         return journal_add_journal_head(bh);
635 }
636
637 /*
638  * Management for journal control blocks: functions to create and
639  * destroy journal_t structures, and to initialise and read existing
640  * journal blocks from disk.  */
641
642 /* First: create and setup a journal_t object in memory.  We initialise
643  * very few fields yet: that has to wait until we have created the
644  * journal structures from from scratch, or loaded them from disk. */
645
646 static journal_t * journal_init_common (void)
647 {
648         journal_t *journal;
649         int err;
650
651         journal = jbd_kmalloc(sizeof(*journal), GFP_KERNEL);
652         if (!journal)
653                 goto fail;
654         memset(journal, 0, sizeof(*journal));
655
656         init_waitqueue_head(&journal->j_wait_transaction_locked);
657         init_waitqueue_head(&journal->j_wait_logspace);
658         init_waitqueue_head(&journal->j_wait_done_commit);
659         init_waitqueue_head(&journal->j_wait_checkpoint);
660         init_waitqueue_head(&journal->j_wait_commit);
661         init_waitqueue_head(&journal->j_wait_updates);
662         init_MUTEX(&journal->j_barrier);
663         init_MUTEX(&journal->j_checkpoint_sem);
664         spin_lock_init(&journal->j_revoke_lock);
665         spin_lock_init(&journal->j_list_lock);
666         spin_lock_init(&journal->j_state_lock);
667
668         journal->j_commit_interval = (HZ * JBD_DEFAULT_MAX_COMMIT_AGE);
669
670         /* The journal is marked for error until we succeed with recovery! */
671         journal->j_flags = JFS_ABORT;
672
673         /* Set up a default-sized revoke table for the new mount. */
674         err = journal_init_revoke(journal, JOURNAL_REVOKE_DEFAULT_HASH);
675         if (err) {
676                 kfree(journal);
677                 goto fail;
678         }
679         return journal;
680 fail:
681         return NULL;
682 }
683
684 /* journal_init_dev and journal_init_inode:
685  *
686  * Create a journal structure assigned some fixed set of disk blocks to
687  * the journal.  We don't actually touch those disk blocks yet, but we
688  * need to set up all of the mapping information to tell the journaling
689  * system where the journal blocks are.
690  *
691  */
692
693 /**
694  *  journal_t * journal_init_dev() - creates an initialises a journal structure
695  *  @bdev: Block device on which to create the journal
696  *  @fs_dev: Device which hold journalled filesystem for this journal.
697  *  @start: Block nr Start of journal.
698  *  @len:  Lenght of the journal in blocks.
699  *  @blocksize: blocksize of journalling device
700  *  @returns: a newly created journal_t *
701  *  
702  *  journal_init_dev creates a journal which maps a fixed contiguous
703  *  range of blocks on an arbitrary block device.
704  * 
705  */
706 journal_t * journal_init_dev(struct block_device *bdev,
707                         struct block_device *fs_dev,
708                         int start, int len, int blocksize)
709 {
710         journal_t *journal = journal_init_common();
711         struct buffer_head *bh;
712         int n;
713
714         if (!journal)
715                 return NULL;
716
717         journal->j_dev = bdev;
718         journal->j_fs_dev = fs_dev;
719         journal->j_blk_offset = start;
720         journal->j_maxlen = len;
721         journal->j_blocksize = blocksize;
722
723         bh = __getblk(journal->j_dev, start, journal->j_blocksize);
724         J_ASSERT(bh != NULL);
725         journal->j_sb_buffer = bh;
726         journal->j_superblock = (journal_superblock_t *)bh->b_data;
727
728         /* journal descriptor can store up to n blocks -bzzz */
729         n = journal->j_blocksize / sizeof(journal_block_tag_t);
730         journal->j_wbufsize = n;
731         journal->j_wbuf = kmalloc(n * sizeof(struct buffer_head*), GFP_KERNEL);
732         if (!journal->j_wbuf) {
733                 printk(KERN_ERR "%s: Cant allocate bhs for commit thread\n",
734                         __FUNCTION__);
735                 kfree(journal);
736                 journal = NULL;
737         }
738
739         return journal;
740 }
741  
742 /** 
743  *  journal_t * journal_init_inode () - creates a journal which maps to a inode.
744  *  @inode: An inode to create the journal in
745  *  
746  * journal_init_inode creates a journal which maps an on-disk inode as
747  * the journal.  The inode must exist already, must support bmap() and
748  * must have all data blocks preallocated.
749  */
750 journal_t * journal_init_inode (struct inode *inode)
751 {
752         struct buffer_head *bh;
753         journal_t *journal = journal_init_common();
754         int err;
755         int n;
756         unsigned long blocknr;
757
758         if (!journal)
759                 return NULL;
760
761         journal->j_dev = journal->j_fs_dev = inode->i_sb->s_bdev;
762         journal->j_inode = inode;
763         jbd_debug(1,
764                   "journal %p: inode %s/%ld, size %Ld, bits %d, blksize %ld\n",
765                   journal, inode->i_sb->s_id, inode->i_ino, 
766                   (long long) inode->i_size,
767                   inode->i_sb->s_blocksize_bits, inode->i_sb->s_blocksize);
768
769         journal->j_maxlen = inode->i_size >> inode->i_sb->s_blocksize_bits;
770         journal->j_blocksize = inode->i_sb->s_blocksize;
771
772         /* journal descriptor can store up to n blocks -bzzz */
773         n = journal->j_blocksize / sizeof(journal_block_tag_t);
774         journal->j_wbufsize = n;
775         journal->j_wbuf = kmalloc(n * sizeof(struct buffer_head*), GFP_KERNEL);
776         if (!journal->j_wbuf) {
777                 printk(KERN_ERR "%s: Cant allocate bhs for commit thread\n",
778                         __FUNCTION__);
779                 kfree(journal);
780                 return NULL;
781         }
782
783         err = journal_bmap(journal, 0, &blocknr);
784         /* If that failed, give up */
785         if (err) {
786                 printk(KERN_ERR "%s: Cannnot locate journal superblock\n",
787                        __FUNCTION__);
788                 kfree(journal);
789                 return NULL;
790         }
791
792         bh = __getblk(journal->j_dev, blocknr, journal->j_blocksize);
793         J_ASSERT(bh != NULL);
794         journal->j_sb_buffer = bh;
795         journal->j_superblock = (journal_superblock_t *)bh->b_data;
796
797         return journal;
798 }
799
800 /* 
801  * If the journal init or create aborts, we need to mark the journal
802  * superblock as being NULL to prevent the journal destroy from writing
803  * back a bogus superblock. 
804  */
805 static void journal_fail_superblock (journal_t *journal)
806 {
807         struct buffer_head *bh = journal->j_sb_buffer;
808         brelse(bh);
809         journal->j_sb_buffer = NULL;
810 }
811
812 /*
813  * Given a journal_t structure, initialise the various fields for
814  * startup of a new journaling session.  We use this both when creating
815  * a journal, and after recovering an old journal to reset it for
816  * subsequent use.
817  */
818
819 static int journal_reset(journal_t *journal)
820 {
821         journal_superblock_t *sb = journal->j_superblock;
822         unsigned int first, last;
823
824         first = be32_to_cpu(sb->s_first);
825         last = be32_to_cpu(sb->s_maxlen);
826
827         journal->j_first = first;
828         journal->j_last = last;
829
830         journal->j_head = first;
831         journal->j_tail = first;
832         journal->j_free = last - first;
833
834         journal->j_tail_sequence = journal->j_transaction_sequence;
835         journal->j_commit_sequence = journal->j_transaction_sequence - 1;
836         journal->j_commit_request = journal->j_commit_sequence;
837
838         journal->j_max_transaction_buffers = journal->j_maxlen / 4;
839
840         /* Add the dynamic fields and write it to disk. */
841         journal_update_superblock(journal, 1);
842         journal_start_thread(journal);
843         return 0;
844 }
845
846 /** 
847  * int journal_create() - Initialise the new journal file
848  * @journal: Journal to create. This structure must have been initialised
849  * 
850  * Given a journal_t structure which tells us which disk blocks we can
851  * use, create a new journal superblock and initialise all of the
852  * journal fields from scratch.  
853  **/
854 int journal_create(journal_t *journal)
855 {
856         unsigned long blocknr;
857         struct buffer_head *bh;
858         journal_superblock_t *sb;
859         int i, err;
860
861         if (journal->j_maxlen < JFS_MIN_JOURNAL_BLOCKS) {
862                 printk (KERN_ERR "Journal length (%d blocks) too short.\n",
863                         journal->j_maxlen);
864                 journal_fail_superblock(journal);
865                 return -EINVAL;
866         }
867
868         if (journal->j_inode == NULL) {
869                 /*
870                  * We don't know what block to start at!
871                  */
872                 printk(KERN_EMERG
873                        "%s: creation of journal on external device!\n",
874                        __FUNCTION__);
875                 BUG();
876         }
877
878         /* Zero out the entire journal on disk.  We cannot afford to
879            have any blocks on disk beginning with JFS_MAGIC_NUMBER. */
880         jbd_debug(1, "JBD: Zeroing out journal blocks...\n");
881         for (i = 0; i < journal->j_maxlen; i++) {
882                 err = journal_bmap(journal, i, &blocknr);
883                 if (err)
884                         return err;
885                 bh = __getblk(journal->j_dev, blocknr, journal->j_blocksize);
886                 lock_buffer(bh);
887                 memset (bh->b_data, 0, journal->j_blocksize);
888                 BUFFER_TRACE(bh, "marking dirty");
889                 mark_buffer_dirty(bh);
890                 BUFFER_TRACE(bh, "marking uptodate");
891                 set_buffer_uptodate(bh);
892                 unlock_buffer(bh);
893                 __brelse(bh);
894         }
895
896         sync_blockdev(journal->j_dev);
897         jbd_debug(1, "JBD: journal cleared.\n");
898
899         /* OK, fill in the initial static fields in the new superblock */
900         sb = journal->j_superblock;
901
902         sb->s_header.h_magic     = cpu_to_be32(JFS_MAGIC_NUMBER);
903         sb->s_header.h_blocktype = cpu_to_be32(JFS_SUPERBLOCK_V2);
904
905         sb->s_blocksize = cpu_to_be32(journal->j_blocksize);
906         sb->s_maxlen    = cpu_to_be32(journal->j_maxlen);
907         sb->s_first     = cpu_to_be32(1);
908
909         journal->j_transaction_sequence = 1;
910
911         journal->j_flags &= ~JFS_ABORT;
912         journal->j_format_version = 2;
913
914         return journal_reset(journal);
915 }
916
917 /** 
918  * void journal_update_superblock() - Update journal sb on disk.
919  * @journal: The journal to update.
920  * @wait: Set to '0' if you don't want to wait for IO completion.
921  *
922  * Update a journal's dynamic superblock fields and write it to disk,
923  * optionally waiting for the IO to complete.
924  */
925 void journal_update_superblock(journal_t *journal, int wait)
926 {
927         journal_superblock_t *sb = journal->j_superblock;
928         struct buffer_head *bh = journal->j_sb_buffer;
929
930         /*
931          * As a special case, if the on-disk copy is already marked as needing
932          * no recovery (s_start == 0) and there are no outstanding transactions
933          * in the filesystem, then we can safely defer the superblock update
934          * until the next commit by setting JFS_FLUSHED.  This avoids
935          * attempting a write to a potential-readonly device.
936          */
937         if (sb->s_start == 0 && journal->j_tail_sequence ==
938                                 journal->j_transaction_sequence) {
939                 jbd_debug(1,"JBD: Skipping superblock update on recovered sb "
940                         "(start %ld, seq %d, errno %d)\n",
941                         journal->j_tail, journal->j_tail_sequence, 
942                         journal->j_errno);
943                 goto out;
944         }
945
946         spin_lock(&journal->j_state_lock);
947         jbd_debug(1,"JBD: updating superblock (start %ld, seq %d, errno %d)\n",
948                   journal->j_tail, journal->j_tail_sequence, journal->j_errno);
949
950         sb->s_sequence = cpu_to_be32(journal->j_tail_sequence);
951         sb->s_start    = cpu_to_be32(journal->j_tail);
952         sb->s_errno    = cpu_to_be32(journal->j_errno);
953         spin_unlock(&journal->j_state_lock);
954
955         BUFFER_TRACE(bh, "marking dirty");
956         mark_buffer_dirty(bh);
957         if (wait)
958                 sync_dirty_buffer(bh);
959         else
960                 ll_rw_block(SWRITE, 1, &bh);
961
962 out:
963         /* If we have just flushed the log (by marking s_start==0), then
964          * any future commit will have to be careful to update the
965          * superblock again to re-record the true start of the log. */
966
967         spin_lock(&journal->j_state_lock);
968         if (sb->s_start)
969                 journal->j_flags &= ~JFS_FLUSHED;
970         else
971                 journal->j_flags |= JFS_FLUSHED;
972         spin_unlock(&journal->j_state_lock);
973 }
974
975 /*
976  * Read the superblock for a given journal, performing initial
977  * validation of the format.
978  */
979
980 static int journal_get_superblock(journal_t *journal)
981 {
982         struct buffer_head *bh;
983         journal_superblock_t *sb;
984         int err = -EIO;
985
986         bh = journal->j_sb_buffer;
987
988         J_ASSERT(bh != NULL);
989         if (!buffer_uptodate(bh)) {
990                 ll_rw_block(READ, 1, &bh);
991                 wait_on_buffer(bh);
992                 if (!buffer_uptodate(bh)) {
993                         printk (KERN_ERR
994                                 "JBD: IO error reading journal superblock\n");
995                         goto out;
996                 }
997         }
998
999         sb = journal->j_superblock;
1000
1001         err = -EINVAL;
1002
1003         if (sb->s_header.h_magic != cpu_to_be32(JFS_MAGIC_NUMBER) ||
1004             sb->s_blocksize != cpu_to_be32(journal->j_blocksize)) {
1005                 printk(KERN_WARNING "JBD: no valid journal superblock found\n");
1006                 goto out;
1007         }
1008
1009         switch(be32_to_cpu(sb->s_header.h_blocktype)) {
1010         case JFS_SUPERBLOCK_V1:
1011                 journal->j_format_version = 1;
1012                 break;
1013         case JFS_SUPERBLOCK_V2:
1014                 journal->j_format_version = 2;
1015                 break;
1016         default:
1017                 printk(KERN_WARNING "JBD: unrecognised superblock format ID\n");
1018                 goto out;
1019         }
1020
1021         if (be32_to_cpu(sb->s_maxlen) < journal->j_maxlen)
1022                 journal->j_maxlen = be32_to_cpu(sb->s_maxlen);
1023         else if (be32_to_cpu(sb->s_maxlen) > journal->j_maxlen) {
1024                 printk (KERN_WARNING "JBD: journal file too short\n");
1025                 goto out;
1026         }
1027
1028         return 0;
1029
1030 out:
1031         journal_fail_superblock(journal);
1032         return err;
1033 }
1034
1035 /*
1036  * Load the on-disk journal superblock and read the key fields into the
1037  * journal_t.
1038  */
1039
1040 static int load_superblock(journal_t *journal)
1041 {
1042         int err;
1043         journal_superblock_t *sb;
1044
1045         err = journal_get_superblock(journal);
1046         if (err)
1047                 return err;
1048
1049         sb = journal->j_superblock;
1050
1051         journal->j_tail_sequence = be32_to_cpu(sb->s_sequence);
1052         journal->j_tail = be32_to_cpu(sb->s_start);
1053         journal->j_first = be32_to_cpu(sb->s_first);
1054         journal->j_last = be32_to_cpu(sb->s_maxlen);
1055         journal->j_errno = be32_to_cpu(sb->s_errno);
1056
1057         return 0;
1058 }
1059
1060
1061 /**
1062  * int journal_load() - Read journal from disk.
1063  * @journal: Journal to act on.
1064  * 
1065  * Given a journal_t structure which tells us which disk blocks contain
1066  * a journal, read the journal from disk to initialise the in-memory
1067  * structures.
1068  */
1069 int journal_load(journal_t *journal)
1070 {
1071         int err;
1072
1073         err = load_superblock(journal);
1074         if (err)
1075                 return err;
1076
1077         /* If this is a V2 superblock, then we have to check the
1078          * features flags on it. */
1079
1080         if (journal->j_format_version >= 2) {
1081                 journal_superblock_t *sb = journal->j_superblock;
1082
1083                 if ((sb->s_feature_ro_compat &
1084                      ~cpu_to_be32(JFS_KNOWN_ROCOMPAT_FEATURES)) ||
1085                     (sb->s_feature_incompat &
1086                      ~cpu_to_be32(JFS_KNOWN_INCOMPAT_FEATURES))) {
1087                         printk (KERN_WARNING
1088                                 "JBD: Unrecognised features on journal\n");
1089                         return -EINVAL;
1090                 }
1091         }
1092
1093         /* Let the recovery code check whether it needs to recover any
1094          * data from the journal. */
1095         if (journal_recover(journal))
1096                 goto recovery_error;
1097
1098         /* OK, we've finished with the dynamic journal bits:
1099          * reinitialise the dynamic contents of the superblock in memory
1100          * and reset them on disk. */
1101         if (journal_reset(journal))
1102                 goto recovery_error;
1103
1104         journal->j_flags &= ~JFS_ABORT;
1105         journal->j_flags |= JFS_LOADED;
1106         return 0;
1107
1108 recovery_error:
1109         printk (KERN_WARNING "JBD: recovery failed\n");
1110         return -EIO;
1111 }
1112
1113 /**
1114  * void journal_destroy() - Release a journal_t structure.
1115  * @journal: Journal to act on.
1116  *
1117  * Release a journal_t structure once it is no longer in use by the
1118  * journaled object.
1119  */
1120 void journal_destroy(journal_t *journal)
1121 {
1122         /* Wait for the commit thread to wake up and die. */
1123         journal_kill_thread(journal);
1124
1125         /* Force a final log commit */
1126         if (journal->j_running_transaction)
1127                 journal_commit_transaction(journal);
1128
1129         /* Force any old transactions to disk */
1130
1131         /* Totally anal locking here... */
1132         spin_lock(&journal->j_list_lock);
1133         while (journal->j_checkpoint_transactions != NULL) {
1134                 spin_unlock(&journal->j_list_lock);
1135                 log_do_checkpoint(journal);
1136                 spin_lock(&journal->j_list_lock);
1137         }
1138
1139         J_ASSERT(journal->j_running_transaction == NULL);
1140         J_ASSERT(journal->j_committing_transaction == NULL);
1141         J_ASSERT(journal->j_checkpoint_transactions == NULL);
1142         spin_unlock(&journal->j_list_lock);
1143
1144         /* We can now mark the journal as empty. */
1145         journal->j_tail = 0;
1146         journal->j_tail_sequence = ++journal->j_transaction_sequence;
1147         if (journal->j_sb_buffer) {
1148                 journal_update_superblock(journal, 1);
1149                 brelse(journal->j_sb_buffer);
1150         }
1151
1152         if (journal->j_inode)
1153                 iput(journal->j_inode);
1154         if (journal->j_revoke)
1155                 journal_destroy_revoke(journal);
1156         kfree(journal->j_wbuf);
1157         kfree(journal);
1158 }
1159
1160
1161 /**
1162  *int journal_check_used_features () - Check if features specified are used.
1163  * @journal: Journal to check.
1164  * @compat: bitmask of compatible features
1165  * @ro: bitmask of features that force read-only mount
1166  * @incompat: bitmask of incompatible features
1167  * 
1168  * Check whether the journal uses all of a given set of
1169  * features.  Return true (non-zero) if it does. 
1170  **/
1171
1172 int journal_check_used_features (journal_t *journal, unsigned long compat,
1173                                  unsigned long ro, unsigned long incompat)
1174 {
1175         journal_superblock_t *sb;
1176
1177         if (!compat && !ro && !incompat)
1178                 return 1;
1179         if (journal->j_format_version == 1)
1180                 return 0;
1181
1182         sb = journal->j_superblock;
1183
1184         if (((be32_to_cpu(sb->s_feature_compat) & compat) == compat) &&
1185             ((be32_to_cpu(sb->s_feature_ro_compat) & ro) == ro) &&
1186             ((be32_to_cpu(sb->s_feature_incompat) & incompat) == incompat))
1187                 return 1;
1188
1189         return 0;
1190 }
1191
1192 /**
1193  * int journal_check_available_features() - Check feature set in journalling layer
1194  * @journal: Journal to check.
1195  * @compat: bitmask of compatible features
1196  * @ro: bitmask of features that force read-only mount
1197  * @incompat: bitmask of incompatible features
1198  * 
1199  * Check whether the journaling code supports the use of
1200  * all of a given set of features on this journal.  Return true
1201  * (non-zero) if it can. */
1202
1203 int journal_check_available_features (journal_t *journal, unsigned long compat,
1204                                       unsigned long ro, unsigned long incompat)
1205 {
1206         journal_superblock_t *sb;
1207
1208         if (!compat && !ro && !incompat)
1209                 return 1;
1210
1211         sb = journal->j_superblock;
1212
1213         /* We can support any known requested features iff the
1214          * superblock is in version 2.  Otherwise we fail to support any
1215          * extended sb features. */
1216
1217         if (journal->j_format_version != 2)
1218                 return 0;
1219
1220         if ((compat   & JFS_KNOWN_COMPAT_FEATURES) == compat &&
1221             (ro       & JFS_KNOWN_ROCOMPAT_FEATURES) == ro &&
1222             (incompat & JFS_KNOWN_INCOMPAT_FEATURES) == incompat)
1223                 return 1;
1224
1225         return 0;
1226 }
1227
1228 /**
1229  * int journal_set_features () - Mark a given journal feature in the superblock
1230  * @journal: Journal to act on.
1231  * @compat: bitmask of compatible features
1232  * @ro: bitmask of features that force read-only mount
1233  * @incompat: bitmask of incompatible features
1234  *
1235  * Mark a given journal feature as present on the
1236  * superblock.  Returns true if the requested features could be set. 
1237  *
1238  */
1239
1240 int journal_set_features (journal_t *journal, unsigned long compat,
1241                           unsigned long ro, unsigned long incompat)
1242 {
1243         journal_superblock_t *sb;
1244
1245         if (journal_check_used_features(journal, compat, ro, incompat))
1246                 return 1;
1247
1248         if (!journal_check_available_features(journal, compat, ro, incompat))
1249                 return 0;
1250
1251         jbd_debug(1, "Setting new features 0x%lx/0x%lx/0x%lx\n",
1252                   compat, ro, incompat);
1253
1254         sb = journal->j_superblock;
1255
1256         sb->s_feature_compat    |= cpu_to_be32(compat);
1257         sb->s_feature_ro_compat |= cpu_to_be32(ro);
1258         sb->s_feature_incompat  |= cpu_to_be32(incompat);
1259
1260         return 1;
1261 }
1262
1263
1264 /**
1265  * int journal_update_format () - Update on-disk journal structure.
1266  * @journal: Journal to act on.
1267  *
1268  * Given an initialised but unloaded journal struct, poke about in the
1269  * on-disk structure to update it to the most recent supported version.
1270  */
1271 int journal_update_format (journal_t *journal)
1272 {
1273         journal_superblock_t *sb;
1274         int err;
1275
1276         err = journal_get_superblock(journal);
1277         if (err)
1278                 return err;
1279
1280         sb = journal->j_superblock;
1281
1282         switch (be32_to_cpu(sb->s_header.h_blocktype)) {
1283         case JFS_SUPERBLOCK_V2:
1284                 return 0;
1285         case JFS_SUPERBLOCK_V1:
1286                 return journal_convert_superblock_v1(journal, sb);
1287         default:
1288                 break;
1289         }
1290         return -EINVAL;
1291 }
1292
1293 static int journal_convert_superblock_v1(journal_t *journal,
1294                                          journal_superblock_t *sb)
1295 {
1296         int offset, blocksize;
1297         struct buffer_head *bh;
1298
1299         printk(KERN_WARNING
1300                 "JBD: Converting superblock from version 1 to 2.\n");
1301
1302         /* Pre-initialise new fields to zero */
1303         offset = ((char *) &(sb->s_feature_compat)) - ((char *) sb);
1304         blocksize = be32_to_cpu(sb->s_blocksize);
1305         memset(&sb->s_feature_compat, 0, blocksize-offset);
1306
1307         sb->s_nr_users = cpu_to_be32(1);
1308         sb->s_header.h_blocktype = cpu_to_be32(JFS_SUPERBLOCK_V2);
1309         journal->j_format_version = 2;
1310
1311         bh = journal->j_sb_buffer;
1312         BUFFER_TRACE(bh, "marking dirty");
1313         mark_buffer_dirty(bh);
1314         sync_dirty_buffer(bh);
1315         return 0;
1316 }
1317
1318
1319 /**
1320  * int journal_flush () - Flush journal
1321  * @journal: Journal to act on.
1322  * 
1323  * Flush all data for a given journal to disk and empty the journal.
1324  * Filesystems can use this when remounting readonly to ensure that
1325  * recovery does not need to happen on remount.
1326  */
1327
1328 int journal_flush(journal_t *journal)
1329 {
1330         int err = 0;
1331         transaction_t *transaction = NULL;
1332         unsigned long old_tail;
1333
1334         spin_lock(&journal->j_state_lock);
1335
1336         /* Force everything buffered to the log... */
1337         if (journal->j_running_transaction) {
1338                 transaction = journal->j_running_transaction;
1339                 __log_start_commit(journal, transaction->t_tid);
1340         } else if (journal->j_committing_transaction)
1341                 transaction = journal->j_committing_transaction;
1342
1343         /* Wait for the log commit to complete... */
1344         if (transaction) {
1345                 tid_t tid = transaction->t_tid;
1346
1347                 spin_unlock(&journal->j_state_lock);
1348                 log_wait_commit(journal, tid);
1349         } else {
1350                 spin_unlock(&journal->j_state_lock);
1351         }
1352
1353         /* ...and flush everything in the log out to disk. */
1354         spin_lock(&journal->j_list_lock);
1355         while (!err && journal->j_checkpoint_transactions != NULL) {
1356                 spin_unlock(&journal->j_list_lock);
1357                 err = log_do_checkpoint(journal);
1358                 spin_lock(&journal->j_list_lock);
1359         }
1360         spin_unlock(&journal->j_list_lock);
1361         cleanup_journal_tail(journal);
1362
1363         /* Finally, mark the journal as really needing no recovery.
1364          * This sets s_start==0 in the underlying superblock, which is
1365          * the magic code for a fully-recovered superblock.  Any future
1366          * commits of data to the journal will restore the current
1367          * s_start value. */
1368         spin_lock(&journal->j_state_lock);
1369         old_tail = journal->j_tail;
1370         journal->j_tail = 0;
1371         spin_unlock(&journal->j_state_lock);
1372         journal_update_superblock(journal, 1);
1373         spin_lock(&journal->j_state_lock);
1374         journal->j_tail = old_tail;
1375
1376         J_ASSERT(!journal->j_running_transaction);
1377         J_ASSERT(!journal->j_committing_transaction);
1378         J_ASSERT(!journal->j_checkpoint_transactions);
1379         J_ASSERT(journal->j_head == journal->j_tail);
1380         J_ASSERT(journal->j_tail_sequence == journal->j_transaction_sequence);
1381         spin_unlock(&journal->j_state_lock);
1382         return err;
1383 }
1384
1385 /**
1386  * int journal_wipe() - Wipe journal contents
1387  * @journal: Journal to act on.
1388  * @write: flag (see below)
1389  * 
1390  * Wipe out all of the contents of a journal, safely.  This will produce
1391  * a warning if the journal contains any valid recovery information.
1392  * Must be called between journal_init_*() and journal_load().
1393  *
1394  * If 'write' is non-zero, then we wipe out the journal on disk; otherwise
1395  * we merely suppress recovery.
1396  */
1397
1398 int journal_wipe(journal_t *journal, int write)
1399 {
1400         journal_superblock_t *sb;
1401         int err = 0;
1402
1403         J_ASSERT (!(journal->j_flags & JFS_LOADED));
1404
1405         err = load_superblock(journal);
1406         if (err)
1407                 return err;
1408
1409         sb = journal->j_superblock;
1410
1411         if (!journal->j_tail)
1412                 goto no_recovery;
1413
1414         printk (KERN_WARNING "JBD: %s recovery information on journal\n",
1415                 write ? "Clearing" : "Ignoring");
1416
1417         err = journal_skip_recovery(journal);
1418         if (write)
1419                 journal_update_superblock(journal, 1);
1420
1421  no_recovery:
1422         return err;
1423 }
1424
1425 /*
1426  * journal_dev_name: format a character string to describe on what
1427  * device this journal is present.
1428  */
1429
1430 static const char *journal_dev_name(journal_t *journal, char *buffer)
1431 {
1432         struct block_device *bdev;
1433
1434         if (journal->j_inode)
1435                 bdev = journal->j_inode->i_sb->s_bdev;
1436         else
1437                 bdev = journal->j_dev;
1438
1439         return bdevname(bdev, buffer);
1440 }
1441
1442 /*
1443  * Journal abort has very specific semantics, which we describe
1444  * for journal abort. 
1445  *
1446  * Two internal function, which provide abort to te jbd layer
1447  * itself are here.
1448  */
1449
1450 /*
1451  * Quick version for internal journal use (doesn't lock the journal).
1452  * Aborts hard --- we mark the abort as occurred, but do _nothing_ else,
1453  * and don't attempt to make any other journal updates.
1454  */
1455 void __journal_abort_hard(journal_t *journal)
1456 {
1457         transaction_t *transaction;
1458         char b[BDEVNAME_SIZE];
1459
1460         if (journal->j_flags & JFS_ABORT)
1461                 return;
1462
1463         printk(KERN_ERR "Aborting journal on device %s.\n",
1464                 journal_dev_name(journal, b));
1465
1466         spin_lock(&journal->j_state_lock);
1467         journal->j_flags |= JFS_ABORT;
1468         transaction = journal->j_running_transaction;
1469         if (transaction)
1470                 __log_start_commit(journal, transaction->t_tid);
1471         spin_unlock(&journal->j_state_lock);
1472 }
1473
1474 /* Soft abort: record the abort error status in the journal superblock,
1475  * but don't do any other IO. */
1476 static void __journal_abort_soft (journal_t *journal, int errno)
1477 {
1478         if (journal->j_flags & JFS_ABORT)
1479                 return;
1480
1481         if (!journal->j_errno)
1482                 journal->j_errno = errno;
1483
1484         __journal_abort_hard(journal);
1485
1486         if (errno)
1487                 journal_update_superblock(journal, 1);
1488 }
1489
1490 /**
1491  * void journal_abort () - Shutdown the journal immediately.
1492  * @journal: the journal to shutdown.
1493  * @errno:   an error number to record in the journal indicating
1494  *           the reason for the shutdown.
1495  *
1496  * Perform a complete, immediate shutdown of the ENTIRE
1497  * journal (not of a single transaction).  This operation cannot be
1498  * undone without closing and reopening the journal.
1499  *           
1500  * The journal_abort function is intended to support higher level error
1501  * recovery mechanisms such as the ext2/ext3 remount-readonly error
1502  * mode.
1503  *
1504  * Journal abort has very specific semantics.  Any existing dirty,
1505  * unjournaled buffers in the main filesystem will still be written to
1506  * disk by bdflush, but the journaling mechanism will be suspended
1507  * immediately and no further transaction commits will be honoured.
1508  *
1509  * Any dirty, journaled buffers will be written back to disk without
1510  * hitting the journal.  Atomicity cannot be guaranteed on an aborted
1511  * filesystem, but we _do_ attempt to leave as much data as possible
1512  * behind for fsck to use for cleanup.
1513  *
1514  * Any attempt to get a new transaction handle on a journal which is in
1515  * ABORT state will just result in an -EROFS error return.  A
1516  * journal_stop on an existing handle will return -EIO if we have
1517  * entered abort state during the update.
1518  *
1519  * Recursive transactions are not disturbed by journal abort until the
1520  * final journal_stop, which will receive the -EIO error.
1521  *
1522  * Finally, the journal_abort call allows the caller to supply an errno
1523  * which will be recorded (if possible) in the journal superblock.  This
1524  * allows a client to record failure conditions in the middle of a
1525  * transaction without having to complete the transaction to record the
1526  * failure to disk.  ext3_error, for example, now uses this
1527  * functionality.
1528  *
1529  * Errors which originate from within the journaling layer will NOT
1530  * supply an errno; a null errno implies that absolutely no further
1531  * writes are done to the journal (unless there are any already in
1532  * progress).
1533  * 
1534  */
1535
1536 void journal_abort(journal_t *journal, int errno)
1537 {
1538         __journal_abort_soft(journal, errno);
1539 }
1540
1541 /** 
1542  * int journal_errno () - returns the journal's error state.
1543  * @journal: journal to examine.
1544  *
1545  * This is the errno numbet set with journal_abort(), the last
1546  * time the journal was mounted - if the journal was stopped
1547  * without calling abort this will be 0.
1548  *
1549  * If the journal has been aborted on this mount time -EROFS will
1550  * be returned.
1551  */
1552 int journal_errno(journal_t *journal)
1553 {
1554         int err;
1555
1556         spin_lock(&journal->j_state_lock);
1557         if (journal->j_flags & JFS_ABORT)
1558                 err = -EROFS;
1559         else
1560                 err = journal->j_errno;
1561         spin_unlock(&journal->j_state_lock);
1562         return err;
1563 }
1564
1565 /** 
1566  * int journal_clear_err () - clears the journal's error state
1567  * @journal: journal to act on.
1568  *
1569  * An error must be cleared or Acked to take a FS out of readonly
1570  * mode.
1571  */
1572 int journal_clear_err(journal_t *journal)
1573 {
1574         int err = 0;
1575
1576         spin_lock(&journal->j_state_lock);
1577         if (journal->j_flags & JFS_ABORT)
1578                 err = -EROFS;
1579         else
1580                 journal->j_errno = 0;
1581         spin_unlock(&journal->j_state_lock);
1582         return err;
1583 }
1584
1585 /** 
1586  * void journal_ack_err() - Ack journal err.
1587  * @journal: journal to act on.
1588  *
1589  * An error must be cleared or Acked to take a FS out of readonly
1590  * mode.
1591  */
1592 void journal_ack_err(journal_t *journal)
1593 {
1594         spin_lock(&journal->j_state_lock);
1595         if (journal->j_errno)
1596                 journal->j_flags |= JFS_ACK_ERR;
1597         spin_unlock(&journal->j_state_lock);
1598 }
1599
1600 int journal_blocks_per_page(struct inode *inode)
1601 {
1602         return 1 << (PAGE_CACHE_SHIFT - inode->i_sb->s_blocksize_bits);
1603 }
1604
1605 /*
1606  * Simple support for retrying memory allocations.  Introduced to help to
1607  * debug different VM deadlock avoidance strategies. 
1608  */
1609 void * __jbd_kmalloc (const char *where, size_t size, gfp_t flags, int retry)
1610 {
1611         return kmalloc(size, flags | (retry ? __GFP_NOFAIL : 0));
1612 }
1613
1614 /*
1615  * Journal_head storage management
1616  */
1617 static kmem_cache_t *journal_head_cache;
1618 #ifdef CONFIG_JBD_DEBUG
1619 static atomic_t nr_journal_heads = ATOMIC_INIT(0);
1620 #endif
1621
1622 static int journal_init_journal_head_cache(void)
1623 {
1624         int retval;
1625
1626         J_ASSERT(journal_head_cache == 0);
1627         journal_head_cache = kmem_cache_create("journal_head",
1628                                 sizeof(struct journal_head),
1629                                 0,              /* offset */
1630                                 0,              /* flags */
1631                                 NULL,           /* ctor */
1632                                 NULL);          /* dtor */
1633         retval = 0;
1634         if (journal_head_cache == 0) {
1635                 retval = -ENOMEM;
1636                 printk(KERN_EMERG "JBD: no memory for journal_head cache\n");
1637         }
1638         return retval;
1639 }
1640
1641 static void journal_destroy_journal_head_cache(void)
1642 {
1643         J_ASSERT(journal_head_cache != NULL);
1644         kmem_cache_destroy(journal_head_cache);
1645         journal_head_cache = NULL;
1646 }
1647
1648 /*
1649  * journal_head splicing and dicing
1650  */
1651 static struct journal_head *journal_alloc_journal_head(void)
1652 {
1653         struct journal_head *ret;
1654         static unsigned long last_warning;
1655
1656 #ifdef CONFIG_JBD_DEBUG
1657         atomic_inc(&nr_journal_heads);
1658 #endif
1659         ret = kmem_cache_alloc(journal_head_cache, GFP_NOFS);
1660         if (ret == 0) {
1661                 jbd_debug(1, "out of memory for journal_head\n");
1662                 if (time_after(jiffies, last_warning + 5*HZ)) {
1663                         printk(KERN_NOTICE "ENOMEM in %s, retrying.\n",
1664                                __FUNCTION__);
1665                         last_warning = jiffies;
1666                 }
1667                 while (ret == 0) {
1668                         yield();
1669                         ret = kmem_cache_alloc(journal_head_cache, GFP_NOFS);
1670                 }
1671         }
1672         return ret;
1673 }
1674
1675 static void journal_free_journal_head(struct journal_head *jh)
1676 {
1677 #ifdef CONFIG_JBD_DEBUG
1678         atomic_dec(&nr_journal_heads);
1679         memset(jh, 0x5b, sizeof(*jh));
1680 #endif
1681         kmem_cache_free(journal_head_cache, jh);
1682 }
1683
1684 /*
1685  * A journal_head is attached to a buffer_head whenever JBD has an
1686  * interest in the buffer.
1687  *
1688  * Whenever a buffer has an attached journal_head, its ->b_state:BH_JBD bit
1689  * is set.  This bit is tested in core kernel code where we need to take
1690  * JBD-specific actions.  Testing the zeroness of ->b_private is not reliable
1691  * there.
1692  *
1693  * When a buffer has its BH_JBD bit set, its ->b_count is elevated by one.
1694  *
1695  * When a buffer has its BH_JBD bit set it is immune from being released by
1696  * core kernel code, mainly via ->b_count.
1697  *
1698  * A journal_head may be detached from its buffer_head when the journal_head's
1699  * b_transaction, b_cp_transaction and b_next_transaction pointers are NULL.
1700  * Various places in JBD call journal_remove_journal_head() to indicate that the
1701  * journal_head can be dropped if needed.
1702  *
1703  * Various places in the kernel want to attach a journal_head to a buffer_head
1704  * _before_ attaching the journal_head to a transaction.  To protect the
1705  * journal_head in this situation, journal_add_journal_head elevates the
1706  * journal_head's b_jcount refcount by one.  The caller must call
1707  * journal_put_journal_head() to undo this.
1708  *
1709  * So the typical usage would be:
1710  *
1711  *      (Attach a journal_head if needed.  Increments b_jcount)
1712  *      struct journal_head *jh = journal_add_journal_head(bh);
1713  *      ...
1714  *      jh->b_transaction = xxx;
1715  *      journal_put_journal_head(jh);
1716  *
1717  * Now, the journal_head's b_jcount is zero, but it is safe from being released
1718  * because it has a non-zero b_transaction.
1719  */
1720
1721 /*
1722  * Give a buffer_head a journal_head.
1723  *
1724  * Doesn't need the journal lock.
1725  * May sleep.
1726  */
1727 struct journal_head *journal_add_journal_head(struct buffer_head *bh)
1728 {
1729         struct journal_head *jh;
1730         struct journal_head *new_jh = NULL;
1731
1732 repeat:
1733         if (!buffer_jbd(bh)) {
1734                 new_jh = journal_alloc_journal_head();
1735                 memset(new_jh, 0, sizeof(*new_jh));
1736         }
1737
1738         jbd_lock_bh_journal_head(bh);
1739         if (buffer_jbd(bh)) {
1740                 jh = bh2jh(bh);
1741         } else {
1742                 J_ASSERT_BH(bh,
1743                         (atomic_read(&bh->b_count) > 0) ||
1744                         (bh->b_page && bh->b_page->mapping));
1745
1746                 if (!new_jh) {
1747                         jbd_unlock_bh_journal_head(bh);
1748                         goto repeat;
1749                 }
1750
1751                 jh = new_jh;
1752                 new_jh = NULL;          /* We consumed it */
1753                 set_buffer_jbd(bh);
1754                 bh->b_private = jh;
1755                 jh->b_bh = bh;
1756                 get_bh(bh);
1757                 BUFFER_TRACE(bh, "added journal_head");
1758         }
1759         jh->b_jcount++;
1760         jbd_unlock_bh_journal_head(bh);
1761         if (new_jh)
1762                 journal_free_journal_head(new_jh);
1763         return bh->b_private;
1764 }
1765
1766 /*
1767  * Grab a ref against this buffer_head's journal_head.  If it ended up not
1768  * having a journal_head, return NULL
1769  */
1770 struct journal_head *journal_grab_journal_head(struct buffer_head *bh)
1771 {
1772         struct journal_head *jh = NULL;
1773
1774         jbd_lock_bh_journal_head(bh);
1775         if (buffer_jbd(bh)) {
1776                 jh = bh2jh(bh);
1777                 jh->b_jcount++;
1778         }
1779         jbd_unlock_bh_journal_head(bh);
1780         return jh;
1781 }
1782
1783 static void __journal_remove_journal_head(struct buffer_head *bh)
1784 {
1785         struct journal_head *jh = bh2jh(bh);
1786
1787         J_ASSERT_JH(jh, jh->b_jcount >= 0);
1788
1789         get_bh(bh);
1790         if (jh->b_jcount == 0) {
1791                 if (jh->b_transaction == NULL &&
1792                                 jh->b_next_transaction == NULL &&
1793                                 jh->b_cp_transaction == NULL) {
1794                         J_ASSERT_JH(jh, jh->b_jlist == BJ_None);
1795                         J_ASSERT_BH(bh, buffer_jbd(bh));
1796                         J_ASSERT_BH(bh, jh2bh(jh) == bh);
1797                         BUFFER_TRACE(bh, "remove journal_head");
1798                         if (jh->b_frozen_data) {
1799                                 printk(KERN_WARNING "%s: freeing "
1800                                                 "b_frozen_data\n",
1801                                                 __FUNCTION__);
1802                                 kfree(jh->b_frozen_data);
1803                         }
1804                         if (jh->b_committed_data) {
1805                                 printk(KERN_WARNING "%s: freeing "
1806                                                 "b_committed_data\n",
1807                                                 __FUNCTION__);
1808                                 kfree(jh->b_committed_data);
1809                         }
1810                         bh->b_private = NULL;
1811                         jh->b_bh = NULL;        /* debug, really */
1812                         clear_buffer_jbd(bh);
1813                         __brelse(bh);
1814                         journal_free_journal_head(jh);
1815                 } else {
1816                         BUFFER_TRACE(bh, "journal_head was locked");
1817                 }
1818         }
1819 }
1820
1821 /*
1822  * journal_remove_journal_head(): if the buffer isn't attached to a transaction
1823  * and has a zero b_jcount then remove and release its journal_head.   If we did
1824  * see that the buffer is not used by any transaction we also "logically"
1825  * decrement ->b_count.
1826  *
1827  * We in fact take an additional increment on ->b_count as a convenience,
1828  * because the caller usually wants to do additional things with the bh
1829  * after calling here.
1830  * The caller of journal_remove_journal_head() *must* run __brelse(bh) at some
1831  * time.  Once the caller has run __brelse(), the buffer is eligible for
1832  * reaping by try_to_free_buffers().
1833  */
1834 void journal_remove_journal_head(struct buffer_head *bh)
1835 {
1836         jbd_lock_bh_journal_head(bh);
1837         __journal_remove_journal_head(bh);
1838         jbd_unlock_bh_journal_head(bh);
1839 }
1840
1841 /*
1842  * Drop a reference on the passed journal_head.  If it fell to zero then try to
1843  * release the journal_head from the buffer_head.
1844  */
1845 void journal_put_journal_head(struct journal_head *jh)
1846 {
1847         struct buffer_head *bh = jh2bh(jh);
1848
1849         jbd_lock_bh_journal_head(bh);
1850         J_ASSERT_JH(jh, jh->b_jcount > 0);
1851         --jh->b_jcount;
1852         if (!jh->b_jcount && !jh->b_transaction) {
1853                 __journal_remove_journal_head(bh);
1854                 __brelse(bh);
1855         }
1856         jbd_unlock_bh_journal_head(bh);
1857 }
1858
1859 /*
1860  * /proc tunables
1861  */
1862 #if defined(CONFIG_JBD_DEBUG)
1863 int journal_enable_debug;
1864 EXPORT_SYMBOL(journal_enable_debug);
1865 #endif
1866
1867 #if defined(CONFIG_JBD_DEBUG) && defined(CONFIG_PROC_FS)
1868
1869 static struct proc_dir_entry *proc_jbd_debug;
1870
1871 static int read_jbd_debug(char *page, char **start, off_t off,
1872                           int count, int *eof, void *data)
1873 {
1874         int ret;
1875
1876         ret = sprintf(page + off, "%d\n", journal_enable_debug);
1877         *eof = 1;
1878         return ret;
1879 }
1880
1881 static int write_jbd_debug(struct file *file, const char __user *buffer,
1882                            unsigned long count, void *data)
1883 {
1884         char buf[32];
1885
1886         if (count > ARRAY_SIZE(buf) - 1)
1887                 count = ARRAY_SIZE(buf) - 1;
1888         if (copy_from_user(buf, buffer, count))
1889                 return -EFAULT;
1890         buf[ARRAY_SIZE(buf) - 1] = '\0';
1891         journal_enable_debug = simple_strtoul(buf, NULL, 10);
1892         return count;
1893 }
1894
1895 #define JBD_PROC_NAME "sys/fs/jbd-debug"
1896
1897 static void __init create_jbd_proc_entry(void)
1898 {
1899         proc_jbd_debug = create_proc_entry(JBD_PROC_NAME, 0644, NULL);
1900         if (proc_jbd_debug) {
1901                 /* Why is this so hard? */
1902                 proc_jbd_debug->read_proc = read_jbd_debug;
1903                 proc_jbd_debug->write_proc = write_jbd_debug;
1904         }
1905 }
1906
1907 static void __exit remove_jbd_proc_entry(void)
1908 {
1909         if (proc_jbd_debug)
1910                 remove_proc_entry(JBD_PROC_NAME, NULL);
1911 }
1912
1913 #else
1914
1915 #define create_jbd_proc_entry() do {} while (0)
1916 #define remove_jbd_proc_entry() do {} while (0)
1917
1918 #endif
1919
1920 kmem_cache_t *jbd_handle_cache;
1921
1922 static int __init journal_init_handle_cache(void)
1923 {
1924         jbd_handle_cache = kmem_cache_create("journal_handle",
1925                                 sizeof(handle_t),
1926                                 0,              /* offset */
1927                                 0,              /* flags */
1928                                 NULL,           /* ctor */
1929                                 NULL);          /* dtor */
1930         if (jbd_handle_cache == NULL) {
1931                 printk(KERN_EMERG "JBD: failed to create handle cache\n");
1932                 return -ENOMEM;
1933         }
1934         return 0;
1935 }
1936
1937 static void journal_destroy_handle_cache(void)
1938 {
1939         if (jbd_handle_cache)
1940                 kmem_cache_destroy(jbd_handle_cache);
1941 }
1942
1943 /*
1944  * Module startup and shutdown
1945  */
1946
1947 static int __init journal_init_caches(void)
1948 {
1949         int ret;
1950
1951         ret = journal_init_revoke_caches();
1952         if (ret == 0)
1953                 ret = journal_init_journal_head_cache();
1954         if (ret == 0)
1955                 ret = journal_init_handle_cache();
1956         return ret;
1957 }
1958
1959 static void journal_destroy_caches(void)
1960 {
1961         journal_destroy_revoke_caches();
1962         journal_destroy_journal_head_cache();
1963         journal_destroy_handle_cache();
1964 }
1965
1966 static int __init journal_init(void)
1967 {
1968         int ret;
1969
1970 /* Static check for data structure consistency.  There's no code
1971  * invoked --- we'll just get a linker failure if things aren't right.
1972  */
1973         extern void journal_bad_superblock_size(void);
1974         if (sizeof(struct journal_superblock_s) != 1024)
1975                 journal_bad_superblock_size();
1976
1977
1978         ret = journal_init_caches();
1979         if (ret != 0)
1980                 journal_destroy_caches();
1981         create_jbd_proc_entry();
1982         return ret;
1983 }
1984
1985 static void __exit journal_exit(void)
1986 {
1987 #ifdef CONFIG_JBD_DEBUG
1988         int n = atomic_read(&nr_journal_heads);
1989         if (n)
1990                 printk(KERN_EMERG "JBD: leaked %d journal_heads!\n", n);
1991 #endif
1992         remove_jbd_proc_entry();
1993         journal_destroy_caches();
1994 }
1995
1996 MODULE_LICENSE("GPL");
1997 module_init(journal_init);
1998 module_exit(journal_exit);
1999