65c290470664f05f7be9a04e1c95b1b70d3a4c81
[linux-3.10.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
39 #include <asm/uaccess.h>
40 #include <asm/page.h>
41
42 EXPORT_SYMBOL(journal_start);
43 EXPORT_SYMBOL(journal_restart);
44 EXPORT_SYMBOL(journal_extend);
45 EXPORT_SYMBOL(journal_stop);
46 EXPORT_SYMBOL(journal_lock_updates);
47 EXPORT_SYMBOL(journal_unlock_updates);
48 EXPORT_SYMBOL(journal_get_write_access);
49 EXPORT_SYMBOL(journal_get_create_access);
50 EXPORT_SYMBOL(journal_get_undo_access);
51 EXPORT_SYMBOL(journal_dirty_data);
52 EXPORT_SYMBOL(journal_dirty_metadata);
53 EXPORT_SYMBOL(journal_release_buffer);
54 EXPORT_SYMBOL(journal_forget);
55 #if 0
56 EXPORT_SYMBOL(journal_sync_buffer);
57 #endif
58 EXPORT_SYMBOL(journal_flush);
59 EXPORT_SYMBOL(journal_revoke);
60
61 EXPORT_SYMBOL(journal_init_dev);
62 EXPORT_SYMBOL(journal_init_inode);
63 EXPORT_SYMBOL(journal_update_format);
64 EXPORT_SYMBOL(journal_check_used_features);
65 EXPORT_SYMBOL(journal_check_available_features);
66 EXPORT_SYMBOL(journal_set_features);
67 EXPORT_SYMBOL(journal_create);
68 EXPORT_SYMBOL(journal_load);
69 EXPORT_SYMBOL(journal_destroy);
70 EXPORT_SYMBOL(journal_update_superblock);
71 EXPORT_SYMBOL(journal_abort);
72 EXPORT_SYMBOL(journal_errno);
73 EXPORT_SYMBOL(journal_ack_err);
74 EXPORT_SYMBOL(journal_clear_err);
75 EXPORT_SYMBOL(log_wait_commit);
76 EXPORT_SYMBOL(journal_start_commit);
77 EXPORT_SYMBOL(journal_force_commit_nested);
78 EXPORT_SYMBOL(journal_wipe);
79 EXPORT_SYMBOL(journal_blocks_per_page);
80 EXPORT_SYMBOL(journal_invalidatepage);
81 EXPORT_SYMBOL(journal_try_to_free_buffers);
82 EXPORT_SYMBOL(journal_force_commit);
83
84 static int journal_convert_superblock_v1(journal_t *, journal_superblock_t *);
85 static void __journal_abort_soft (journal_t *journal, int errno);
86
87 /*
88  * Helper function used to manage commit timeouts
89  */
90
91 static void commit_timeout(unsigned long __data)
92 {
93         struct task_struct * p = (struct task_struct *) __data;
94
95         wake_up_process(p);
96 }
97
98 /*
99  * kjournald: The main thread function used to manage a logging device
100  * journal.
101  *
102  * This kernel thread is responsible for two things:
103  *
104  * 1) COMMIT:  Every so often we need to commit the current state of the
105  *    filesystem to disk.  The journal thread is responsible for writing
106  *    all of the metadata buffers to disk.
107  *
108  * 2) CHECKPOINT: We cannot reuse a used section of the log file until all
109  *    of the data in that part of the log has been rewritten elsewhere on
110  *    the disk.  Flushing these old buffers to reclaim space in the log is
111  *    known as checkpointing, and this thread is responsible for that job.
112  */
113
114 static int kjournald(void *arg)
115 {
116         journal_t *journal = arg;
117         transaction_t *transaction;
118
119         /*
120          * Set up an interval timer which can be used to trigger a commit wakeup
121          * after the commit interval expires
122          */
123         setup_timer(&journal->j_commit_timer, commit_timeout,
124                         (unsigned long)current);
125
126         /* Record that the journal thread is running */
127         journal->j_task = current;
128         wake_up(&journal->j_wait_done_commit);
129
130         printk(KERN_INFO "kjournald starting.  Commit interval %ld seconds\n",
131                         journal->j_commit_interval / HZ);
132
133         /*
134          * And now, wait forever for commit wakeup events.
135          */
136         spin_lock(&journal->j_state_lock);
137
138 loop:
139         if (journal->j_flags & JFS_UNMOUNT)
140                 goto end_loop;
141
142         jbd_debug(1, "commit_sequence=%d, commit_request=%d\n",
143                 journal->j_commit_sequence, journal->j_commit_request);
144
145         if (journal->j_commit_sequence != journal->j_commit_request) {
146                 jbd_debug(1, "OK, requests differ\n");
147                 spin_unlock(&journal->j_state_lock);
148                 del_timer_sync(&journal->j_commit_timer);
149                 journal_commit_transaction(journal);
150                 spin_lock(&journal->j_state_lock);
151                 goto loop;
152         }
153
154         wake_up(&journal->j_wait_done_commit);
155         if (freezing(current)) {
156                 /*
157                  * The simpler the better. Flushing journal isn't a
158                  * good idea, because that depends on threads that may
159                  * be already stopped.
160                  */
161                 jbd_debug(1, "Now suspending kjournald\n");
162                 spin_unlock(&journal->j_state_lock);
163                 refrigerator();
164                 spin_lock(&journal->j_state_lock);
165         } else {
166                 /*
167                  * We assume on resume that commits are already there,
168                  * so we don't sleep
169                  */
170                 DEFINE_WAIT(wait);
171                 int should_sleep = 1;
172
173                 prepare_to_wait(&journal->j_wait_commit, &wait,
174                                 TASK_INTERRUPTIBLE);
175                 if (journal->j_commit_sequence != journal->j_commit_request)
176                         should_sleep = 0;
177                 transaction = journal->j_running_transaction;
178                 if (transaction && time_after_eq(jiffies,
179                                                 transaction->t_expires))
180                         should_sleep = 0;
181                 if (journal->j_flags & JFS_UNMOUNT)
182                         should_sleep = 0;
183                 if (should_sleep) {
184                         spin_unlock(&journal->j_state_lock);
185                         schedule();
186                         spin_lock(&journal->j_state_lock);
187                 }
188                 finish_wait(&journal->j_wait_commit, &wait);
189         }
190
191         jbd_debug(1, "kjournald wakes\n");
192
193         /*
194          * Were we woken up by a commit wakeup event?
195          */
196         transaction = journal->j_running_transaction;
197         if (transaction && time_after_eq(jiffies, transaction->t_expires)) {
198                 journal->j_commit_request = transaction->t_tid;
199                 jbd_debug(1, "woke because of timeout\n");
200         }
201         goto loop;
202
203 end_loop:
204         spin_unlock(&journal->j_state_lock);
205         del_timer_sync(&journal->j_commit_timer);
206         journal->j_task = NULL;
207         wake_up(&journal->j_wait_done_commit);
208         jbd_debug(1, "Journal thread exiting.\n");
209         return 0;
210 }
211
212 static int journal_start_thread(journal_t *journal)
213 {
214         struct task_struct *t;
215
216         t = kthread_run(kjournald, journal, "kjournald");
217         if (IS_ERR(t))
218                 return PTR_ERR(t);
219
220         wait_event(journal->j_wait_done_commit, journal->j_task != 0);
221         return 0;
222 }
223
224 static void journal_kill_thread(journal_t *journal)
225 {
226         spin_lock(&journal->j_state_lock);
227         journal->j_flags |= JFS_UNMOUNT;
228
229         while (journal->j_task) {
230                 wake_up(&journal->j_wait_commit);
231                 spin_unlock(&journal->j_state_lock);
232                 wait_event(journal->j_wait_done_commit, journal->j_task == 0);
233                 spin_lock(&journal->j_state_lock);
234         }
235         spin_unlock(&journal->j_state_lock);
236 }
237
238 /*
239  * journal_write_metadata_buffer: write a metadata buffer to the journal.
240  *
241  * Writes a metadata buffer to a given disk block.  The actual IO is not
242  * performed but a new buffer_head is constructed which labels the data
243  * to be written with the correct destination disk block.
244  *
245  * Any magic-number escaping which needs to be done will cause a
246  * copy-out here.  If the buffer happens to start with the
247  * JFS_MAGIC_NUMBER, then we can't write it to the log directly: the
248  * magic number is only written to the log for descripter blocks.  In
249  * this case, we copy the data and replace the first word with 0, and we
250  * return a result code which indicates that this buffer needs to be
251  * marked as an escaped buffer in the corresponding log descriptor
252  * block.  The missing word can then be restored when the block is read
253  * during recovery.
254  *
255  * If the source buffer has already been modified by a new transaction
256  * since we took the last commit snapshot, we use the frozen copy of
257  * that data for IO.  If we end up using the existing buffer_head's data
258  * for the write, then we *have* to lock the buffer to prevent anyone
259  * else from using and possibly modifying it while the IO is in
260  * progress.
261  *
262  * The function returns a pointer to the buffer_heads to be used for IO.
263  *
264  * We assume that the journal has already been locked in this function.
265  *
266  * Return value:
267  *  <0: Error
268  * >=0: Finished OK
269  *
270  * On success:
271  * Bit 0 set == escape performed on the data
272  * Bit 1 set == buffer copy-out performed (kfree the data after IO)
273  */
274
275 int journal_write_metadata_buffer(transaction_t *transaction,
276                                   struct journal_head  *jh_in,
277                                   struct journal_head **jh_out,
278                                   unsigned long blocknr)
279 {
280         int need_copy_out = 0;
281         int done_copy_out = 0;
282         int do_escape = 0;
283         char *mapped_data;
284         struct buffer_head *new_bh;
285         struct journal_head *new_jh;
286         struct page *new_page;
287         unsigned int new_offset;
288         struct buffer_head *bh_in = jh2bh(jh_in);
289
290         /*
291          * The buffer really shouldn't be locked: only the current committing
292          * transaction is allowed to write it, so nobody else is allowed
293          * to do any IO.
294          *
295          * akpm: except if we're journalling data, and write() output is
296          * also part of a shared mapping, and another thread has
297          * decided to launch a writepage() against this buffer.
298          */
299         J_ASSERT_BH(bh_in, buffer_jbddirty(bh_in));
300
301         new_bh = alloc_buffer_head(GFP_NOFS|__GFP_NOFAIL);
302
303         /*
304          * If a new transaction has already done a buffer copy-out, then
305          * we use that version of the data for the commit.
306          */
307         jbd_lock_bh_state(bh_in);
308 repeat:
309         if (jh_in->b_frozen_data) {
310                 done_copy_out = 1;
311                 new_page = virt_to_page(jh_in->b_frozen_data);
312                 new_offset = offset_in_page(jh_in->b_frozen_data);
313         } else {
314                 new_page = jh2bh(jh_in)->b_page;
315                 new_offset = offset_in_page(jh2bh(jh_in)->b_data);
316         }
317
318         mapped_data = kmap_atomic(new_page, KM_USER0);
319         /*
320          * Check for escaping
321          */
322         if (*((__be32 *)(mapped_data + new_offset)) ==
323                                 cpu_to_be32(JFS_MAGIC_NUMBER)) {
324                 need_copy_out = 1;
325                 do_escape = 1;
326         }
327         kunmap_atomic(mapped_data, KM_USER0);
328
329         /*
330          * Do we need to do a data copy?
331          */
332         if (need_copy_out && !done_copy_out) {
333                 char *tmp;
334
335                 jbd_unlock_bh_state(bh_in);
336                 tmp = jbd_alloc(bh_in->b_size, GFP_NOFS);
337                 jbd_lock_bh_state(bh_in);
338                 if (jh_in->b_frozen_data) {
339                         jbd_free(tmp, bh_in->b_size);
340                         goto repeat;
341                 }
342
343                 jh_in->b_frozen_data = tmp;
344                 mapped_data = kmap_atomic(new_page, KM_USER0);
345                 memcpy(tmp, mapped_data + new_offset, jh2bh(jh_in)->b_size);
346                 kunmap_atomic(mapped_data, KM_USER0);
347
348                 new_page = virt_to_page(tmp);
349                 new_offset = offset_in_page(tmp);
350                 done_copy_out = 1;
351         }
352
353         /*
354          * Did we need to do an escaping?  Now we've done all the
355          * copying, we can finally do so.
356          */
357         if (do_escape) {
358                 mapped_data = kmap_atomic(new_page, KM_USER0);
359                 *((unsigned int *)(mapped_data + new_offset)) = 0;
360                 kunmap_atomic(mapped_data, KM_USER0);
361         }
362
363         /* keep subsequent assertions sane */
364         new_bh->b_state = 0;
365         init_buffer(new_bh, NULL, NULL);
366         atomic_set(&new_bh->b_count, 1);
367         jbd_unlock_bh_state(bh_in);
368
369         new_jh = journal_add_journal_head(new_bh);      /* This sleeps */
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         journal_file_buffer(jh_in, transaction, BJ_Shadow);
388         JBUFFER_TRACE(new_jh, "file as BJ_IO");
389         journal_file_buffer(new_jh, transaction, BJ_IO);
390
391         return do_escape | (done_copy_out << 1);
392 }
393
394 /*
395  * Allocation code for the journal file.  Manage the space left in the
396  * journal, so that we can begin checkpointing when appropriate.
397  */
398
399 /*
400  * __log_space_left: Return the number of free blocks left in the journal.
401  *
402  * Called with the journal already locked.
403  *
404  * Called under j_state_lock
405  */
406
407 int __log_space_left(journal_t *journal)
408 {
409         int left = journal->j_free;
410
411         assert_spin_locked(&journal->j_state_lock);
412
413         /*
414          * Be pessimistic here about the number of those free blocks which
415          * might be required for log descriptor control blocks.
416          */
417
418 #define MIN_LOG_RESERVED_BLOCKS 32 /* Allow for rounding errors */
419
420         left -= MIN_LOG_RESERVED_BLOCKS;
421
422         if (left <= 0)
423                 return 0;
424         left -= (left >> 3);
425         return left;
426 }
427
428 /*
429  * Called under j_state_lock.  Returns true if a transaction was started.
430  */
431 int __log_start_commit(journal_t *journal, tid_t target)
432 {
433         /*
434          * Are we already doing a recent enough commit?
435          */
436         if (!tid_geq(journal->j_commit_request, target)) {
437                 /*
438                  * We want a new commit: OK, mark the request and wakup the
439                  * commit thread.  We do _not_ do the commit ourselves.
440                  */
441
442                 journal->j_commit_request = target;
443                 jbd_debug(1, "JBD: requesting commit %d/%d\n",
444                           journal->j_commit_request,
445                           journal->j_commit_sequence);
446                 wake_up(&journal->j_wait_commit);
447                 return 1;
448         }
449         return 0;
450 }
451
452 int log_start_commit(journal_t *journal, tid_t tid)
453 {
454         int ret;
455
456         spin_lock(&journal->j_state_lock);
457         ret = __log_start_commit(journal, tid);
458         spin_unlock(&journal->j_state_lock);
459         return ret;
460 }
461
462 /*
463  * Force and wait upon a commit if the calling process is not within
464  * transaction.  This is used for forcing out undo-protected data which contains
465  * bitmaps, when the fs is running out of space.
466  *
467  * We can only force the running transaction if we don't have an active handle;
468  * otherwise, we will deadlock.
469  *
470  * Returns true if a transaction was started.
471  */
472 int journal_force_commit_nested(journal_t *journal)
473 {
474         transaction_t *transaction = NULL;
475         tid_t tid;
476
477         spin_lock(&journal->j_state_lock);
478         if (journal->j_running_transaction && !current->journal_info) {
479                 transaction = journal->j_running_transaction;
480                 __log_start_commit(journal, transaction->t_tid);
481         } else if (journal->j_committing_transaction)
482                 transaction = journal->j_committing_transaction;
483
484         if (!transaction) {
485                 spin_unlock(&journal->j_state_lock);
486                 return 0;       /* Nothing to retry */
487         }
488
489         tid = transaction->t_tid;
490         spin_unlock(&journal->j_state_lock);
491         log_wait_commit(journal, tid);
492         return 1;
493 }
494
495 /*
496  * Start a commit of the current running transaction (if any).  Returns true
497  * if a transaction was started, and fills its tid in at *ptid
498  */
499 int journal_start_commit(journal_t *journal, tid_t *ptid)
500 {
501         int ret = 0;
502
503         spin_lock(&journal->j_state_lock);
504         if (journal->j_running_transaction) {
505                 tid_t tid = journal->j_running_transaction->t_tid;
506
507                 ret = __log_start_commit(journal, tid);
508                 if (ret && ptid)
509                         *ptid = tid;
510         } else if (journal->j_committing_transaction && ptid) {
511                 /*
512                  * If ext3_write_super() recently started a commit, then we
513                  * have to wait for completion of that transaction
514                  */
515                 *ptid = journal->j_committing_transaction->t_tid;
516                 ret = 1;
517         }
518         spin_unlock(&journal->j_state_lock);
519         return ret;
520 }
521
522 /*
523  * Wait for a specified commit to complete.
524  * The caller may not hold the journal lock.
525  */
526 int log_wait_commit(journal_t *journal, tid_t tid)
527 {
528         int err = 0;
529
530 #ifdef CONFIG_JBD_DEBUG
531         spin_lock(&journal->j_state_lock);
532         if (!tid_geq(journal->j_commit_request, tid)) {
533                 printk(KERN_EMERG
534                        "%s: error: j_commit_request=%d, tid=%d\n",
535                        __FUNCTION__, journal->j_commit_request, tid);
536         }
537         spin_unlock(&journal->j_state_lock);
538 #endif
539         spin_lock(&journal->j_state_lock);
540         while (tid_gt(tid, journal->j_commit_sequence)) {
541                 jbd_debug(1, "JBD: want %d, j_commit_sequence=%d\n",
542                                   tid, journal->j_commit_sequence);
543                 wake_up(&journal->j_wait_commit);
544                 spin_unlock(&journal->j_state_lock);
545                 wait_event(journal->j_wait_done_commit,
546                                 !tid_gt(tid, journal->j_commit_sequence));
547                 spin_lock(&journal->j_state_lock);
548         }
549         spin_unlock(&journal->j_state_lock);
550
551         if (unlikely(is_journal_aborted(journal))) {
552                 printk(KERN_EMERG "journal commit I/O error\n");
553                 err = -EIO;
554         }
555         return err;
556 }
557
558 /*
559  * Log buffer allocation routines:
560  */
561
562 int journal_next_log_block(journal_t *journal, unsigned long *retp)
563 {
564         unsigned long blocknr;
565
566         spin_lock(&journal->j_state_lock);
567         J_ASSERT(journal->j_free > 1);
568
569         blocknr = journal->j_head;
570         journal->j_head++;
571         journal->j_free--;
572         if (journal->j_head == journal->j_last)
573                 journal->j_head = journal->j_first;
574         spin_unlock(&journal->j_state_lock);
575         return journal_bmap(journal, blocknr, retp);
576 }
577
578 /*
579  * Conversion of logical to physical block numbers for the journal
580  *
581  * On external journals the journal blocks are identity-mapped, so
582  * this is a no-op.  If needed, we can use j_blk_offset - everything is
583  * ready.
584  */
585 int journal_bmap(journal_t *journal, unsigned long blocknr,
586                  unsigned long *retp)
587 {
588         int err = 0;
589         unsigned long ret;
590
591         if (journal->j_inode) {
592                 ret = bmap(journal->j_inode, blocknr);
593                 if (ret)
594                         *retp = ret;
595                 else {
596                         char b[BDEVNAME_SIZE];
597
598                         printk(KERN_ALERT "%s: journal block not found "
599                                         "at offset %lu on %s\n",
600                                 __FUNCTION__,
601                                 blocknr,
602                                 bdevname(journal->j_dev, b));
603                         err = -EIO;
604                         __journal_abort_soft(journal, err);
605                 }
606         } else {
607                 *retp = blocknr; /* +journal->j_blk_offset */
608         }
609         return err;
610 }
611
612 /*
613  * We play buffer_head aliasing tricks to write data/metadata blocks to
614  * the journal without copying their contents, but for journal
615  * descriptor blocks we do need to generate bona fide buffers.
616  *
617  * After the caller of journal_get_descriptor_buffer() has finished modifying
618  * the buffer's contents they really should run flush_dcache_page(bh->b_page).
619  * But we don't bother doing that, so there will be coherency problems with
620  * mmaps of blockdevs which hold live JBD-controlled filesystems.
621  */
622 struct journal_head *journal_get_descriptor_buffer(journal_t *journal)
623 {
624         struct buffer_head *bh;
625         unsigned long blocknr;
626         int err;
627
628         err = journal_next_log_block(journal, &blocknr);
629
630         if (err)
631                 return NULL;
632
633         bh = __getblk(journal->j_dev, blocknr, journal->j_blocksize);
634         lock_buffer(bh);
635         memset(bh->b_data, 0, journal->j_blocksize);
636         set_buffer_uptodate(bh);
637         unlock_buffer(bh);
638         BUFFER_TRACE(bh, "return this buffer");
639         return journal_add_journal_head(bh);
640 }
641
642 /*
643  * Management for journal control blocks: functions to create and
644  * destroy journal_t structures, and to initialise and read existing
645  * journal blocks from disk.  */
646
647 /* First: create and setup a journal_t object in memory.  We initialise
648  * very few fields yet: that has to wait until we have created the
649  * journal structures from from scratch, or loaded them from disk. */
650
651 static journal_t * journal_init_common (void)
652 {
653         journal_t *journal;
654         int err;
655
656         journal = kmalloc(sizeof(*journal), GFP_KERNEL);
657         if (!journal)
658                 goto fail;
659         memset(journal, 0, sizeof(*journal));
660
661         init_waitqueue_head(&journal->j_wait_transaction_locked);
662         init_waitqueue_head(&journal->j_wait_logspace);
663         init_waitqueue_head(&journal->j_wait_done_commit);
664         init_waitqueue_head(&journal->j_wait_checkpoint);
665         init_waitqueue_head(&journal->j_wait_commit);
666         init_waitqueue_head(&journal->j_wait_updates);
667         mutex_init(&journal->j_barrier);
668         mutex_init(&journal->j_checkpoint_mutex);
669         spin_lock_init(&journal->j_revoke_lock);
670         spin_lock_init(&journal->j_list_lock);
671         spin_lock_init(&journal->j_state_lock);
672
673         journal->j_commit_interval = (HZ * JBD_DEFAULT_MAX_COMMIT_AGE);
674
675         /* The journal is marked for error until we succeed with recovery! */
676         journal->j_flags = JFS_ABORT;
677
678         /* Set up a default-sized revoke table for the new mount. */
679         err = journal_init_revoke(journal, JOURNAL_REVOKE_DEFAULT_HASH);
680         if (err) {
681                 kfree(journal);
682                 goto fail;
683         }
684         return journal;
685 fail:
686         return NULL;
687 }
688
689 /* journal_init_dev and journal_init_inode:
690  *
691  * Create a journal structure assigned some fixed set of disk blocks to
692  * the journal.  We don't actually touch those disk blocks yet, but we
693  * need to set up all of the mapping information to tell the journaling
694  * system where the journal blocks are.
695  *
696  */
697
698 /**
699  *  journal_t * journal_init_dev() - creates an initialises a journal structure
700  *  @bdev: Block device on which to create the journal
701  *  @fs_dev: Device which hold journalled filesystem for this journal.
702  *  @start: Block nr Start of journal.
703  *  @len:  Length of the journal in blocks.
704  *  @blocksize: blocksize of journalling device
705  *  @returns: a newly created journal_t *
706  *
707  *  journal_init_dev creates a journal which maps a fixed contiguous
708  *  range of blocks on an arbitrary block device.
709  *
710  */
711 journal_t * journal_init_dev(struct block_device *bdev,
712                         struct block_device *fs_dev,
713                         int start, int len, int blocksize)
714 {
715         journal_t *journal = journal_init_common();
716         struct buffer_head *bh;
717         int n;
718
719         if (!journal)
720                 return NULL;
721
722         /* journal descriptor can store up to n blocks -bzzz */
723         journal->j_blocksize = blocksize;
724         n = journal->j_blocksize / sizeof(journal_block_tag_t);
725         journal->j_wbufsize = n;
726         journal->j_wbuf = kmalloc(n * sizeof(struct buffer_head*), GFP_KERNEL);
727         if (!journal->j_wbuf) {
728                 printk(KERN_ERR "%s: Cant allocate bhs for commit thread\n",
729                         __FUNCTION__);
730                 kfree(journal);
731                 journal = NULL;
732                 goto out;
733         }
734         journal->j_dev = bdev;
735         journal->j_fs_dev = fs_dev;
736         journal->j_blk_offset = start;
737         journal->j_maxlen = len;
738
739         bh = __getblk(journal->j_dev, start, journal->j_blocksize);
740         J_ASSERT(bh != NULL);
741         journal->j_sb_buffer = bh;
742         journal->j_superblock = (journal_superblock_t *)bh->b_data;
743 out:
744         return journal;
745 }
746
747 /**
748  *  journal_t * journal_init_inode () - creates a journal which maps to a inode.
749  *  @inode: An inode to create the journal in
750  *
751  * journal_init_inode creates a journal which maps an on-disk inode as
752  * the journal.  The inode must exist already, must support bmap() and
753  * must have all data blocks preallocated.
754  */
755 journal_t * journal_init_inode (struct inode *inode)
756 {
757         struct buffer_head *bh;
758         journal_t *journal = journal_init_common();
759         int err;
760         int n;
761         unsigned long blocknr;
762
763         if (!journal)
764                 return NULL;
765
766         journal->j_dev = journal->j_fs_dev = inode->i_sb->s_bdev;
767         journal->j_inode = inode;
768         jbd_debug(1,
769                   "journal %p: inode %s/%ld, size %Ld, bits %d, blksize %ld\n",
770                   journal, inode->i_sb->s_id, inode->i_ino,
771                   (long long) inode->i_size,
772                   inode->i_sb->s_blocksize_bits, inode->i_sb->s_blocksize);
773
774         journal->j_maxlen = inode->i_size >> inode->i_sb->s_blocksize_bits;
775         journal->j_blocksize = inode->i_sb->s_blocksize;
776
777         /* journal descriptor can store up to n blocks -bzzz */
778         n = journal->j_blocksize / sizeof(journal_block_tag_t);
779         journal->j_wbufsize = n;
780         journal->j_wbuf = kmalloc(n * sizeof(struct buffer_head*), GFP_KERNEL);
781         if (!journal->j_wbuf) {
782                 printk(KERN_ERR "%s: Cant allocate bhs for commit thread\n",
783                         __FUNCTION__);
784                 kfree(journal);
785                 return NULL;
786         }
787
788         err = journal_bmap(journal, 0, &blocknr);
789         /* If that failed, give up */
790         if (err) {
791                 printk(KERN_ERR "%s: Cannnot locate journal superblock\n",
792                        __FUNCTION__);
793                 kfree(journal);
794                 return NULL;
795         }
796
797         bh = __getblk(journal->j_dev, blocknr, journal->j_blocksize);
798         J_ASSERT(bh != NULL);
799         journal->j_sb_buffer = bh;
800         journal->j_superblock = (journal_superblock_t *)bh->b_data;
801
802         return journal;
803 }
804
805 /*
806  * If the journal init or create aborts, we need to mark the journal
807  * superblock as being NULL to prevent the journal destroy from writing
808  * back a bogus superblock.
809  */
810 static void journal_fail_superblock (journal_t *journal)
811 {
812         struct buffer_head *bh = journal->j_sb_buffer;
813         brelse(bh);
814         journal->j_sb_buffer = NULL;
815 }
816
817 /*
818  * Given a journal_t structure, initialise the various fields for
819  * startup of a new journaling session.  We use this both when creating
820  * a journal, and after recovering an old journal to reset it for
821  * subsequent use.
822  */
823
824 static int journal_reset(journal_t *journal)
825 {
826         journal_superblock_t *sb = journal->j_superblock;
827         unsigned long first, last;
828
829         first = be32_to_cpu(sb->s_first);
830         last = be32_to_cpu(sb->s_maxlen);
831
832         journal->j_first = first;
833         journal->j_last = last;
834
835         journal->j_head = first;
836         journal->j_tail = first;
837         journal->j_free = last - first;
838
839         journal->j_tail_sequence = journal->j_transaction_sequence;
840         journal->j_commit_sequence = journal->j_transaction_sequence - 1;
841         journal->j_commit_request = journal->j_commit_sequence;
842
843         journal->j_max_transaction_buffers = journal->j_maxlen / 4;
844
845         /* Add the dynamic fields and write it to disk. */
846         journal_update_superblock(journal, 1);
847         return journal_start_thread(journal);
848 }
849
850 /**
851  * int journal_create() - Initialise the new journal file
852  * @journal: Journal to create. This structure must have been initialised
853  *
854  * Given a journal_t structure which tells us which disk blocks we can
855  * use, create a new journal superblock and initialise all of the
856  * journal fields from scratch.
857  **/
858 int journal_create(journal_t *journal)
859 {
860         unsigned long blocknr;
861         struct buffer_head *bh;
862         journal_superblock_t *sb;
863         int i, err;
864
865         if (journal->j_maxlen < JFS_MIN_JOURNAL_BLOCKS) {
866                 printk (KERN_ERR "Journal length (%d blocks) too short.\n",
867                         journal->j_maxlen);
868                 journal_fail_superblock(journal);
869                 return -EINVAL;
870         }
871
872         if (journal->j_inode == NULL) {
873                 /*
874                  * We don't know what block to start at!
875                  */
876                 printk(KERN_EMERG
877                        "%s: creation of journal on external device!\n",
878                        __FUNCTION__);
879                 BUG();
880         }
881
882         /* Zero out the entire journal on disk.  We cannot afford to
883            have any blocks on disk beginning with JFS_MAGIC_NUMBER. */
884         jbd_debug(1, "JBD: Zeroing out journal blocks...\n");
885         for (i = 0; i < journal->j_maxlen; i++) {
886                 err = journal_bmap(journal, i, &blocknr);
887                 if (err)
888                         return err;
889                 bh = __getblk(journal->j_dev, blocknr, journal->j_blocksize);
890                 lock_buffer(bh);
891                 memset (bh->b_data, 0, journal->j_blocksize);
892                 BUFFER_TRACE(bh, "marking dirty");
893                 mark_buffer_dirty(bh);
894                 BUFFER_TRACE(bh, "marking uptodate");
895                 set_buffer_uptodate(bh);
896                 unlock_buffer(bh);
897                 __brelse(bh);
898         }
899
900         sync_blockdev(journal->j_dev);
901         jbd_debug(1, "JBD: journal cleared.\n");
902
903         /* OK, fill in the initial static fields in the new superblock */
904         sb = journal->j_superblock;
905
906         sb->s_header.h_magic     = cpu_to_be32(JFS_MAGIC_NUMBER);
907         sb->s_header.h_blocktype = cpu_to_be32(JFS_SUPERBLOCK_V2);
908
909         sb->s_blocksize = cpu_to_be32(journal->j_blocksize);
910         sb->s_maxlen    = cpu_to_be32(journal->j_maxlen);
911         sb->s_first     = cpu_to_be32(1);
912
913         journal->j_transaction_sequence = 1;
914
915         journal->j_flags &= ~JFS_ABORT;
916         journal->j_format_version = 2;
917
918         return journal_reset(journal);
919 }
920
921 /**
922  * void journal_update_superblock() - Update journal sb on disk.
923  * @journal: The journal to update.
924  * @wait: Set to '0' if you don't want to wait for IO completion.
925  *
926  * Update a journal's dynamic superblock fields and write it to disk,
927  * optionally waiting for the IO to complete.
928  */
929 void journal_update_superblock(journal_t *journal, int wait)
930 {
931         journal_superblock_t *sb = journal->j_superblock;
932         struct buffer_head *bh = journal->j_sb_buffer;
933
934         /*
935          * As a special case, if the on-disk copy is already marked as needing
936          * no recovery (s_start == 0) and there are no outstanding transactions
937          * in the filesystem, then we can safely defer the superblock update
938          * until the next commit by setting JFS_FLUSHED.  This avoids
939          * attempting a write to a potential-readonly device.
940          */
941         if (sb->s_start == 0 && journal->j_tail_sequence ==
942                                 journal->j_transaction_sequence) {
943                 jbd_debug(1,"JBD: Skipping superblock update on recovered sb "
944                         "(start %ld, seq %d, errno %d)\n",
945                         journal->j_tail, journal->j_tail_sequence,
946                         journal->j_errno);
947                 goto out;
948         }
949
950         spin_lock(&journal->j_state_lock);
951         jbd_debug(1,"JBD: updating superblock (start %ld, seq %d, errno %d)\n",
952                   journal->j_tail, journal->j_tail_sequence, journal->j_errno);
953
954         sb->s_sequence = cpu_to_be32(journal->j_tail_sequence);
955         sb->s_start    = cpu_to_be32(journal->j_tail);
956         sb->s_errno    = cpu_to_be32(journal->j_errno);
957         spin_unlock(&journal->j_state_lock);
958
959         BUFFER_TRACE(bh, "marking dirty");
960         mark_buffer_dirty(bh);
961         if (wait)
962                 sync_dirty_buffer(bh);
963         else
964                 ll_rw_block(SWRITE, 1, &bh);
965
966 out:
967         /* If we have just flushed the log (by marking s_start==0), then
968          * any future commit will have to be careful to update the
969          * superblock again to re-record the true start of the log. */
970
971         spin_lock(&journal->j_state_lock);
972         if (sb->s_start)
973                 journal->j_flags &= ~JFS_FLUSHED;
974         else
975                 journal->j_flags |= JFS_FLUSHED;
976         spin_unlock(&journal->j_state_lock);
977 }
978
979 /*
980  * Read the superblock for a given journal, performing initial
981  * validation of the format.
982  */
983
984 static int journal_get_superblock(journal_t *journal)
985 {
986         struct buffer_head *bh;
987         journal_superblock_t *sb;
988         int err = -EIO;
989
990         bh = journal->j_sb_buffer;
991
992         J_ASSERT(bh != NULL);
993         if (!buffer_uptodate(bh)) {
994                 ll_rw_block(READ, 1, &bh);
995                 wait_on_buffer(bh);
996                 if (!buffer_uptodate(bh)) {
997                         printk (KERN_ERR
998                                 "JBD: IO error reading journal superblock\n");
999                         goto out;
1000                 }
1001         }
1002
1003         sb = journal->j_superblock;
1004
1005         err = -EINVAL;
1006
1007         if (sb->s_header.h_magic != cpu_to_be32(JFS_MAGIC_NUMBER) ||
1008             sb->s_blocksize != cpu_to_be32(journal->j_blocksize)) {
1009                 printk(KERN_WARNING "JBD: no valid journal superblock found\n");
1010                 goto out;
1011         }
1012
1013         switch(be32_to_cpu(sb->s_header.h_blocktype)) {
1014         case JFS_SUPERBLOCK_V1:
1015                 journal->j_format_version = 1;
1016                 break;
1017         case JFS_SUPERBLOCK_V2:
1018                 journal->j_format_version = 2;
1019                 break;
1020         default:
1021                 printk(KERN_WARNING "JBD: unrecognised superblock format ID\n");
1022                 goto out;
1023         }
1024
1025         if (be32_to_cpu(sb->s_maxlen) < journal->j_maxlen)
1026                 journal->j_maxlen = be32_to_cpu(sb->s_maxlen);
1027         else if (be32_to_cpu(sb->s_maxlen) > journal->j_maxlen) {
1028                 printk (KERN_WARNING "JBD: journal file too short\n");
1029                 goto out;
1030         }
1031
1032         return 0;
1033
1034 out:
1035         journal_fail_superblock(journal);
1036         return err;
1037 }
1038
1039 /*
1040  * Load the on-disk journal superblock and read the key fields into the
1041  * journal_t.
1042  */
1043
1044 static int load_superblock(journal_t *journal)
1045 {
1046         int err;
1047         journal_superblock_t *sb;
1048
1049         err = journal_get_superblock(journal);
1050         if (err)
1051                 return err;
1052
1053         sb = journal->j_superblock;
1054
1055         journal->j_tail_sequence = be32_to_cpu(sb->s_sequence);
1056         journal->j_tail = be32_to_cpu(sb->s_start);
1057         journal->j_first = be32_to_cpu(sb->s_first);
1058         journal->j_last = be32_to_cpu(sb->s_maxlen);
1059         journal->j_errno = be32_to_cpu(sb->s_errno);
1060
1061         return 0;
1062 }
1063
1064
1065 /**
1066  * int journal_load() - Read journal from disk.
1067  * @journal: Journal to act on.
1068  *
1069  * Given a journal_t structure which tells us which disk blocks contain
1070  * a journal, read the journal from disk to initialise the in-memory
1071  * structures.
1072  */
1073 int journal_load(journal_t *journal)
1074 {
1075         int err;
1076         journal_superblock_t *sb;
1077
1078         err = load_superblock(journal);
1079         if (err)
1080                 return err;
1081
1082         sb = journal->j_superblock;
1083         /* If this is a V2 superblock, then we have to check the
1084          * features flags on it. */
1085
1086         if (journal->j_format_version >= 2) {
1087                 if ((sb->s_feature_ro_compat &
1088                      ~cpu_to_be32(JFS_KNOWN_ROCOMPAT_FEATURES)) ||
1089                     (sb->s_feature_incompat &
1090                      ~cpu_to_be32(JFS_KNOWN_INCOMPAT_FEATURES))) {
1091                         printk (KERN_WARNING
1092                                 "JBD: Unrecognised features on journal\n");
1093                         return -EINVAL;
1094                 }
1095         }
1096
1097         /* Let the recovery code check whether it needs to recover any
1098          * data from the journal. */
1099         if (journal_recover(journal))
1100                 goto recovery_error;
1101
1102         /* OK, we've finished with the dynamic journal bits:
1103          * reinitialise the dynamic contents of the superblock in memory
1104          * and reset them on disk. */
1105         if (journal_reset(journal))
1106                 goto recovery_error;
1107
1108         journal->j_flags &= ~JFS_ABORT;
1109         journal->j_flags |= JFS_LOADED;
1110         return 0;
1111
1112 recovery_error:
1113         printk (KERN_WARNING "JBD: recovery failed\n");
1114         return -EIO;
1115 }
1116
1117 /**
1118  * void journal_destroy() - Release a journal_t structure.
1119  * @journal: Journal to act on.
1120  *
1121  * Release a journal_t structure once it is no longer in use by the
1122  * journaled object.
1123  */
1124 void journal_destroy(journal_t *journal)
1125 {
1126         /* Wait for the commit thread to wake up and die. */
1127         journal_kill_thread(journal);
1128
1129         /* Force a final log commit */
1130         if (journal->j_running_transaction)
1131                 journal_commit_transaction(journal);
1132
1133         /* Force any old transactions to disk */
1134
1135         /* Totally anal locking here... */
1136         spin_lock(&journal->j_list_lock);
1137         while (journal->j_checkpoint_transactions != NULL) {
1138                 spin_unlock(&journal->j_list_lock);
1139                 log_do_checkpoint(journal);
1140                 spin_lock(&journal->j_list_lock);
1141         }
1142
1143         J_ASSERT(journal->j_running_transaction == NULL);
1144         J_ASSERT(journal->j_committing_transaction == NULL);
1145         J_ASSERT(journal->j_checkpoint_transactions == NULL);
1146         spin_unlock(&journal->j_list_lock);
1147
1148         /* We can now mark the journal as empty. */
1149         journal->j_tail = 0;
1150         journal->j_tail_sequence = ++journal->j_transaction_sequence;
1151         if (journal->j_sb_buffer) {
1152                 journal_update_superblock(journal, 1);
1153                 brelse(journal->j_sb_buffer);
1154         }
1155
1156         if (journal->j_inode)
1157                 iput(journal->j_inode);
1158         if (journal->j_revoke)
1159                 journal_destroy_revoke(journal);
1160         kfree(journal->j_wbuf);
1161         kfree(journal);
1162 }
1163
1164
1165 /**
1166  *int journal_check_used_features () - Check if features specified are used.
1167  * @journal: Journal to check.
1168  * @compat: bitmask of compatible features
1169  * @ro: bitmask of features that force read-only mount
1170  * @incompat: bitmask of incompatible features
1171  *
1172  * Check whether the journal uses all of a given set of
1173  * features.  Return true (non-zero) if it does.
1174  **/
1175
1176 int journal_check_used_features (journal_t *journal, unsigned long compat,
1177                                  unsigned long ro, unsigned long incompat)
1178 {
1179         journal_superblock_t *sb;
1180
1181         if (!compat && !ro && !incompat)
1182                 return 1;
1183         if (journal->j_format_version == 1)
1184                 return 0;
1185
1186         sb = journal->j_superblock;
1187
1188         if (((be32_to_cpu(sb->s_feature_compat) & compat) == compat) &&
1189             ((be32_to_cpu(sb->s_feature_ro_compat) & ro) == ro) &&
1190             ((be32_to_cpu(sb->s_feature_incompat) & incompat) == incompat))
1191                 return 1;
1192
1193         return 0;
1194 }
1195
1196 /**
1197  * int journal_check_available_features() - Check feature set in journalling layer
1198  * @journal: Journal to check.
1199  * @compat: bitmask of compatible features
1200  * @ro: bitmask of features that force read-only mount
1201  * @incompat: bitmask of incompatible features
1202  *
1203  * Check whether the journaling code supports the use of
1204  * all of a given set of features on this journal.  Return true
1205  * (non-zero) if it can. */
1206
1207 int journal_check_available_features (journal_t *journal, unsigned long compat,
1208                                       unsigned long ro, unsigned long incompat)
1209 {
1210         journal_superblock_t *sb;
1211
1212         if (!compat && !ro && !incompat)
1213                 return 1;
1214
1215         sb = journal->j_superblock;
1216
1217         /* We can support any known requested features iff the
1218          * superblock is in version 2.  Otherwise we fail to support any
1219          * extended sb features. */
1220
1221         if (journal->j_format_version != 2)
1222                 return 0;
1223
1224         if ((compat   & JFS_KNOWN_COMPAT_FEATURES) == compat &&
1225             (ro       & JFS_KNOWN_ROCOMPAT_FEATURES) == ro &&
1226             (incompat & JFS_KNOWN_INCOMPAT_FEATURES) == incompat)
1227                 return 1;
1228
1229         return 0;
1230 }
1231
1232 /**
1233  * int journal_set_features () - Mark a given journal feature in the superblock
1234  * @journal: Journal to act on.
1235  * @compat: bitmask of compatible features
1236  * @ro: bitmask of features that force read-only mount
1237  * @incompat: bitmask of incompatible features
1238  *
1239  * Mark a given journal feature as present on the
1240  * superblock.  Returns true if the requested features could be set.
1241  *
1242  */
1243
1244 int journal_set_features (journal_t *journal, unsigned long compat,
1245                           unsigned long ro, unsigned long incompat)
1246 {
1247         journal_superblock_t *sb;
1248
1249         if (journal_check_used_features(journal, compat, ro, incompat))
1250                 return 1;
1251
1252         if (!journal_check_available_features(journal, compat, ro, incompat))
1253                 return 0;
1254
1255         jbd_debug(1, "Setting new features 0x%lx/0x%lx/0x%lx\n",
1256                   compat, ro, incompat);
1257
1258         sb = journal->j_superblock;
1259
1260         sb->s_feature_compat    |= cpu_to_be32(compat);
1261         sb->s_feature_ro_compat |= cpu_to_be32(ro);
1262         sb->s_feature_incompat  |= cpu_to_be32(incompat);
1263
1264         return 1;
1265 }
1266
1267
1268 /**
1269  * int journal_update_format () - Update on-disk journal structure.
1270  * @journal: Journal to act on.
1271  *
1272  * Given an initialised but unloaded journal struct, poke about in the
1273  * on-disk structure to update it to the most recent supported version.
1274  */
1275 int journal_update_format (journal_t *journal)
1276 {
1277         journal_superblock_t *sb;
1278         int err;
1279
1280         err = journal_get_superblock(journal);
1281         if (err)
1282                 return err;
1283
1284         sb = journal->j_superblock;
1285
1286         switch (be32_to_cpu(sb->s_header.h_blocktype)) {
1287         case JFS_SUPERBLOCK_V2:
1288                 return 0;
1289         case JFS_SUPERBLOCK_V1:
1290                 return journal_convert_superblock_v1(journal, sb);
1291         default:
1292                 break;
1293         }
1294         return -EINVAL;
1295 }
1296
1297 static int journal_convert_superblock_v1(journal_t *journal,
1298                                          journal_superblock_t *sb)
1299 {
1300         int offset, blocksize;
1301         struct buffer_head *bh;
1302
1303         printk(KERN_WARNING
1304                 "JBD: Converting superblock from version 1 to 2.\n");
1305
1306         /* Pre-initialise new fields to zero */
1307         offset = ((char *) &(sb->s_feature_compat)) - ((char *) sb);
1308         blocksize = be32_to_cpu(sb->s_blocksize);
1309         memset(&sb->s_feature_compat, 0, blocksize-offset);
1310
1311         sb->s_nr_users = cpu_to_be32(1);
1312         sb->s_header.h_blocktype = cpu_to_be32(JFS_SUPERBLOCK_V2);
1313         journal->j_format_version = 2;
1314
1315         bh = journal->j_sb_buffer;
1316         BUFFER_TRACE(bh, "marking dirty");
1317         mark_buffer_dirty(bh);
1318         sync_dirty_buffer(bh);
1319         return 0;
1320 }
1321
1322
1323 /**
1324  * int journal_flush () - Flush journal
1325  * @journal: Journal to act on.
1326  *
1327  * Flush all data for a given journal to disk and empty the journal.
1328  * Filesystems can use this when remounting readonly to ensure that
1329  * recovery does not need to happen on remount.
1330  */
1331
1332 int journal_flush(journal_t *journal)
1333 {
1334         int err = 0;
1335         transaction_t *transaction = NULL;
1336         unsigned long old_tail;
1337
1338         spin_lock(&journal->j_state_lock);
1339
1340         /* Force everything buffered to the log... */
1341         if (journal->j_running_transaction) {
1342                 transaction = journal->j_running_transaction;
1343                 __log_start_commit(journal, transaction->t_tid);
1344         } else if (journal->j_committing_transaction)
1345                 transaction = journal->j_committing_transaction;
1346
1347         /* Wait for the log commit to complete... */
1348         if (transaction) {
1349                 tid_t tid = transaction->t_tid;
1350
1351                 spin_unlock(&journal->j_state_lock);
1352                 log_wait_commit(journal, tid);
1353         } else {
1354                 spin_unlock(&journal->j_state_lock);
1355         }
1356
1357         /* ...and flush everything in the log out to disk. */
1358         spin_lock(&journal->j_list_lock);
1359         while (!err && journal->j_checkpoint_transactions != NULL) {
1360                 spin_unlock(&journal->j_list_lock);
1361                 err = log_do_checkpoint(journal);
1362                 spin_lock(&journal->j_list_lock);
1363         }
1364         spin_unlock(&journal->j_list_lock);
1365         cleanup_journal_tail(journal);
1366
1367         /* Finally, mark the journal as really needing no recovery.
1368          * This sets s_start==0 in the underlying superblock, which is
1369          * the magic code for a fully-recovered superblock.  Any future
1370          * commits of data to the journal will restore the current
1371          * s_start value. */
1372         spin_lock(&journal->j_state_lock);
1373         old_tail = journal->j_tail;
1374         journal->j_tail = 0;
1375         spin_unlock(&journal->j_state_lock);
1376         journal_update_superblock(journal, 1);
1377         spin_lock(&journal->j_state_lock);
1378         journal->j_tail = old_tail;
1379
1380         J_ASSERT(!journal->j_running_transaction);
1381         J_ASSERT(!journal->j_committing_transaction);
1382         J_ASSERT(!journal->j_checkpoint_transactions);
1383         J_ASSERT(journal->j_head == journal->j_tail);
1384         J_ASSERT(journal->j_tail_sequence == journal->j_transaction_sequence);
1385         spin_unlock(&journal->j_state_lock);
1386         return err;
1387 }
1388
1389 /**
1390  * int journal_wipe() - Wipe journal contents
1391  * @journal: Journal to act on.
1392  * @write: flag (see below)
1393  *
1394  * Wipe out all of the contents of a journal, safely.  This will produce
1395  * a warning if the journal contains any valid recovery information.
1396  * Must be called between journal_init_*() and journal_load().
1397  *
1398  * If 'write' is non-zero, then we wipe out the journal on disk; otherwise
1399  * we merely suppress recovery.
1400  */
1401
1402 int journal_wipe(journal_t *journal, int write)
1403 {
1404         journal_superblock_t *sb;
1405         int err = 0;
1406
1407         J_ASSERT (!(journal->j_flags & JFS_LOADED));
1408
1409         err = load_superblock(journal);
1410         if (err)
1411                 return err;
1412
1413         sb = journal->j_superblock;
1414
1415         if (!journal->j_tail)
1416                 goto no_recovery;
1417
1418         printk (KERN_WARNING "JBD: %s recovery information on journal\n",
1419                 write ? "Clearing" : "Ignoring");
1420
1421         err = journal_skip_recovery(journal);
1422         if (write)
1423                 journal_update_superblock(journal, 1);
1424
1425  no_recovery:
1426         return err;
1427 }
1428
1429 /*
1430  * journal_dev_name: format a character string to describe on what
1431  * device this journal is present.
1432  */
1433
1434 static const char *journal_dev_name(journal_t *journal, char *buffer)
1435 {
1436         struct block_device *bdev;
1437
1438         if (journal->j_inode)
1439                 bdev = journal->j_inode->i_sb->s_bdev;
1440         else
1441                 bdev = journal->j_dev;
1442
1443         return bdevname(bdev, buffer);
1444 }
1445
1446 /*
1447  * Journal abort has very specific semantics, which we describe
1448  * for journal abort.
1449  *
1450  * Two internal function, which provide abort to te jbd layer
1451  * itself are here.
1452  */
1453
1454 /*
1455  * Quick version for internal journal use (doesn't lock the journal).
1456  * Aborts hard --- we mark the abort as occurred, but do _nothing_ else,
1457  * and don't attempt to make any other journal updates.
1458  */
1459 void __journal_abort_hard(journal_t *journal)
1460 {
1461         transaction_t *transaction;
1462         char b[BDEVNAME_SIZE];
1463
1464         if (journal->j_flags & JFS_ABORT)
1465                 return;
1466
1467         printk(KERN_ERR "Aborting journal on device %s.\n",
1468                 journal_dev_name(journal, b));
1469
1470         spin_lock(&journal->j_state_lock);
1471         journal->j_flags |= JFS_ABORT;
1472         transaction = journal->j_running_transaction;
1473         if (transaction)
1474                 __log_start_commit(journal, transaction->t_tid);
1475         spin_unlock(&journal->j_state_lock);
1476 }
1477
1478 /* Soft abort: record the abort error status in the journal superblock,
1479  * but don't do any other IO. */
1480 static void __journal_abort_soft (journal_t *journal, int errno)
1481 {
1482         if (journal->j_flags & JFS_ABORT)
1483                 return;
1484
1485         if (!journal->j_errno)
1486                 journal->j_errno = errno;
1487
1488         __journal_abort_hard(journal);
1489
1490         if (errno)
1491                 journal_update_superblock(journal, 1);
1492 }
1493
1494 /**
1495  * void journal_abort () - Shutdown the journal immediately.
1496  * @journal: the journal to shutdown.
1497  * @errno:   an error number to record in the journal indicating
1498  *           the reason for the shutdown.
1499  *
1500  * Perform a complete, immediate shutdown of the ENTIRE
1501  * journal (not of a single transaction).  This operation cannot be
1502  * undone without closing and reopening the journal.
1503  *
1504  * The journal_abort function is intended to support higher level error
1505  * recovery mechanisms such as the ext2/ext3 remount-readonly error
1506  * mode.
1507  *
1508  * Journal abort has very specific semantics.  Any existing dirty,
1509  * unjournaled buffers in the main filesystem will still be written to
1510  * disk by bdflush, but the journaling mechanism will be suspended
1511  * immediately and no further transaction commits will be honoured.
1512  *
1513  * Any dirty, journaled buffers will be written back to disk without
1514  * hitting the journal.  Atomicity cannot be guaranteed on an aborted
1515  * filesystem, but we _do_ attempt to leave as much data as possible
1516  * behind for fsck to use for cleanup.
1517  *
1518  * Any attempt to get a new transaction handle on a journal which is in
1519  * ABORT state will just result in an -EROFS error return.  A
1520  * journal_stop on an existing handle will return -EIO if we have
1521  * entered abort state during the update.
1522  *
1523  * Recursive transactions are not disturbed by journal abort until the
1524  * final journal_stop, which will receive the -EIO error.
1525  *
1526  * Finally, the journal_abort call allows the caller to supply an errno
1527  * which will be recorded (if possible) in the journal superblock.  This
1528  * allows a client to record failure conditions in the middle of a
1529  * transaction without having to complete the transaction to record the
1530  * failure to disk.  ext3_error, for example, now uses this
1531  * functionality.
1532  *
1533  * Errors which originate from within the journaling layer will NOT
1534  * supply an errno; a null errno implies that absolutely no further
1535  * writes are done to the journal (unless there are any already in
1536  * progress).
1537  *
1538  */
1539
1540 void journal_abort(journal_t *journal, int errno)
1541 {
1542         __journal_abort_soft(journal, errno);
1543 }
1544
1545 /**
1546  * int journal_errno () - returns the journal's error state.
1547  * @journal: journal to examine.
1548  *
1549  * This is the errno numbet set with journal_abort(), the last
1550  * time the journal was mounted - if the journal was stopped
1551  * without calling abort this will be 0.
1552  *
1553  * If the journal has been aborted on this mount time -EROFS will
1554  * be returned.
1555  */
1556 int journal_errno(journal_t *journal)
1557 {
1558         int err;
1559
1560         spin_lock(&journal->j_state_lock);
1561         if (journal->j_flags & JFS_ABORT)
1562                 err = -EROFS;
1563         else
1564                 err = journal->j_errno;
1565         spin_unlock(&journal->j_state_lock);
1566         return err;
1567 }
1568
1569 /**
1570  * int journal_clear_err () - clears the journal's error state
1571  * @journal: journal to act on.
1572  *
1573  * An error must be cleared or Acked to take a FS out of readonly
1574  * mode.
1575  */
1576 int journal_clear_err(journal_t *journal)
1577 {
1578         int err = 0;
1579
1580         spin_lock(&journal->j_state_lock);
1581         if (journal->j_flags & JFS_ABORT)
1582                 err = -EROFS;
1583         else
1584                 journal->j_errno = 0;
1585         spin_unlock(&journal->j_state_lock);
1586         return err;
1587 }
1588
1589 /**
1590  * void journal_ack_err() - Ack journal err.
1591  * @journal: journal to act on.
1592  *
1593  * An error must be cleared or Acked to take a FS out of readonly
1594  * mode.
1595  */
1596 void journal_ack_err(journal_t *journal)
1597 {
1598         spin_lock(&journal->j_state_lock);
1599         if (journal->j_errno)
1600                 journal->j_flags |= JFS_ACK_ERR;
1601         spin_unlock(&journal->j_state_lock);
1602 }
1603
1604 int journal_blocks_per_page(struct inode *inode)
1605 {
1606         return 1 << (PAGE_CACHE_SHIFT - inode->i_sb->s_blocksize_bits);
1607 }
1608
1609 /*
1610  * Journal_head storage management
1611  */
1612 static struct kmem_cache *journal_head_cache;
1613 #ifdef CONFIG_JBD_DEBUG
1614 static atomic_t nr_journal_heads = ATOMIC_INIT(0);
1615 #endif
1616
1617 static int journal_init_journal_head_cache(void)
1618 {
1619         int retval;
1620
1621         J_ASSERT(journal_head_cache == 0);
1622         journal_head_cache = kmem_cache_create("journal_head",
1623                                 sizeof(struct journal_head),
1624                                 0,              /* offset */
1625                                 SLAB_TEMPORARY, /* flags */
1626                                 NULL);          /* ctor */
1627         retval = 0;
1628         if (journal_head_cache == 0) {
1629                 retval = -ENOMEM;
1630                 printk(KERN_EMERG "JBD: no memory for journal_head cache\n");
1631         }
1632         return retval;
1633 }
1634
1635 static void journal_destroy_journal_head_cache(void)
1636 {
1637         J_ASSERT(journal_head_cache != NULL);
1638         kmem_cache_destroy(journal_head_cache);
1639         journal_head_cache = NULL;
1640 }
1641
1642 /*
1643  * journal_head splicing and dicing
1644  */
1645 static struct journal_head *journal_alloc_journal_head(void)
1646 {
1647         struct journal_head *ret;
1648         static unsigned long last_warning;
1649
1650 #ifdef CONFIG_JBD_DEBUG
1651         atomic_inc(&nr_journal_heads);
1652 #endif
1653         ret = kmem_cache_alloc(journal_head_cache, GFP_NOFS);
1654         if (ret == 0) {
1655                 jbd_debug(1, "out of memory for journal_head\n");
1656                 if (time_after(jiffies, last_warning + 5*HZ)) {
1657                         printk(KERN_NOTICE "ENOMEM in %s, retrying.\n",
1658                                __FUNCTION__);
1659                         last_warning = jiffies;
1660                 }
1661                 while (ret == 0) {
1662                         yield();
1663                         ret = kmem_cache_alloc(journal_head_cache, GFP_NOFS);
1664                 }
1665         }
1666         return ret;
1667 }
1668
1669 static void journal_free_journal_head(struct journal_head *jh)
1670 {
1671 #ifdef CONFIG_JBD_DEBUG
1672         atomic_dec(&nr_journal_heads);
1673         memset(jh, JBD_POISON_FREE, sizeof(*jh));
1674 #endif
1675         kmem_cache_free(journal_head_cache, jh);
1676 }
1677
1678 /*
1679  * A journal_head is attached to a buffer_head whenever JBD has an
1680  * interest in the buffer.
1681  *
1682  * Whenever a buffer has an attached journal_head, its ->b_state:BH_JBD bit
1683  * is set.  This bit is tested in core kernel code where we need to take
1684  * JBD-specific actions.  Testing the zeroness of ->b_private is not reliable
1685  * there.
1686  *
1687  * When a buffer has its BH_JBD bit set, its ->b_count is elevated by one.
1688  *
1689  * When a buffer has its BH_JBD bit set it is immune from being released by
1690  * core kernel code, mainly via ->b_count.
1691  *
1692  * A journal_head may be detached from its buffer_head when the journal_head's
1693  * b_transaction, b_cp_transaction and b_next_transaction pointers are NULL.
1694  * Various places in JBD call journal_remove_journal_head() to indicate that the
1695  * journal_head can be dropped if needed.
1696  *
1697  * Various places in the kernel want to attach a journal_head to a buffer_head
1698  * _before_ attaching the journal_head to a transaction.  To protect the
1699  * journal_head in this situation, journal_add_journal_head elevates the
1700  * journal_head's b_jcount refcount by one.  The caller must call
1701  * journal_put_journal_head() to undo this.
1702  *
1703  * So the typical usage would be:
1704  *
1705  *      (Attach a journal_head if needed.  Increments b_jcount)
1706  *      struct journal_head *jh = journal_add_journal_head(bh);
1707  *      ...
1708  *      jh->b_transaction = xxx;
1709  *      journal_put_journal_head(jh);
1710  *
1711  * Now, the journal_head's b_jcount is zero, but it is safe from being released
1712  * because it has a non-zero b_transaction.
1713  */
1714
1715 /*
1716  * Give a buffer_head a journal_head.
1717  *
1718  * Doesn't need the journal lock.
1719  * May sleep.
1720  */
1721 struct journal_head *journal_add_journal_head(struct buffer_head *bh)
1722 {
1723         struct journal_head *jh;
1724         struct journal_head *new_jh = NULL;
1725
1726 repeat:
1727         if (!buffer_jbd(bh)) {
1728                 new_jh = journal_alloc_journal_head();
1729                 memset(new_jh, 0, sizeof(*new_jh));
1730         }
1731
1732         jbd_lock_bh_journal_head(bh);
1733         if (buffer_jbd(bh)) {
1734                 jh = bh2jh(bh);
1735         } else {
1736                 J_ASSERT_BH(bh,
1737                         (atomic_read(&bh->b_count) > 0) ||
1738                         (bh->b_page && bh->b_page->mapping));
1739
1740                 if (!new_jh) {
1741                         jbd_unlock_bh_journal_head(bh);
1742                         goto repeat;
1743                 }
1744
1745                 jh = new_jh;
1746                 new_jh = NULL;          /* We consumed it */
1747                 set_buffer_jbd(bh);
1748                 bh->b_private = jh;
1749                 jh->b_bh = bh;
1750                 get_bh(bh);
1751                 BUFFER_TRACE(bh, "added journal_head");
1752         }
1753         jh->b_jcount++;
1754         jbd_unlock_bh_journal_head(bh);
1755         if (new_jh)
1756                 journal_free_journal_head(new_jh);
1757         return bh->b_private;
1758 }
1759
1760 /*
1761  * Grab a ref against this buffer_head's journal_head.  If it ended up not
1762  * having a journal_head, return NULL
1763  */
1764 struct journal_head *journal_grab_journal_head(struct buffer_head *bh)
1765 {
1766         struct journal_head *jh = NULL;
1767
1768         jbd_lock_bh_journal_head(bh);
1769         if (buffer_jbd(bh)) {
1770                 jh = bh2jh(bh);
1771                 jh->b_jcount++;
1772         }
1773         jbd_unlock_bh_journal_head(bh);
1774         return jh;
1775 }
1776
1777 static void __journal_remove_journal_head(struct buffer_head *bh)
1778 {
1779         struct journal_head *jh = bh2jh(bh);
1780
1781         J_ASSERT_JH(jh, jh->b_jcount >= 0);
1782
1783         get_bh(bh);
1784         if (jh->b_jcount == 0) {
1785                 if (jh->b_transaction == NULL &&
1786                                 jh->b_next_transaction == NULL &&
1787                                 jh->b_cp_transaction == NULL) {
1788                         J_ASSERT_JH(jh, jh->b_jlist == BJ_None);
1789                         J_ASSERT_BH(bh, buffer_jbd(bh));
1790                         J_ASSERT_BH(bh, jh2bh(jh) == bh);
1791                         BUFFER_TRACE(bh, "remove journal_head");
1792                         if (jh->b_frozen_data) {
1793                                 printk(KERN_WARNING "%s: freeing "
1794                                                 "b_frozen_data\n",
1795                                                 __FUNCTION__);
1796                                 jbd_free(jh->b_frozen_data, bh->b_size);
1797                         }
1798                         if (jh->b_committed_data) {
1799                                 printk(KERN_WARNING "%s: freeing "
1800                                                 "b_committed_data\n",
1801                                                 __FUNCTION__);
1802                                 jbd_free(jh->b_committed_data, bh->b_size);
1803                         }
1804                         bh->b_private = NULL;
1805                         jh->b_bh = NULL;        /* debug, really */
1806                         clear_buffer_jbd(bh);
1807                         __brelse(bh);
1808                         journal_free_journal_head(jh);
1809                 } else {
1810                         BUFFER_TRACE(bh, "journal_head was locked");
1811                 }
1812         }
1813 }
1814
1815 /*
1816  * journal_remove_journal_head(): if the buffer isn't attached to a transaction
1817  * and has a zero b_jcount then remove and release its journal_head.   If we did
1818  * see that the buffer is not used by any transaction we also "logically"
1819  * decrement ->b_count.
1820  *
1821  * We in fact take an additional increment on ->b_count as a convenience,
1822  * because the caller usually wants to do additional things with the bh
1823  * after calling here.
1824  * The caller of journal_remove_journal_head() *must* run __brelse(bh) at some
1825  * time.  Once the caller has run __brelse(), the buffer is eligible for
1826  * reaping by try_to_free_buffers().
1827  */
1828 void journal_remove_journal_head(struct buffer_head *bh)
1829 {
1830         jbd_lock_bh_journal_head(bh);
1831         __journal_remove_journal_head(bh);
1832         jbd_unlock_bh_journal_head(bh);
1833 }
1834
1835 /*
1836  * Drop a reference on the passed journal_head.  If it fell to zero then try to
1837  * release the journal_head from the buffer_head.
1838  */
1839 void journal_put_journal_head(struct journal_head *jh)
1840 {
1841         struct buffer_head *bh = jh2bh(jh);
1842
1843         jbd_lock_bh_journal_head(bh);
1844         J_ASSERT_JH(jh, jh->b_jcount > 0);
1845         --jh->b_jcount;
1846         if (!jh->b_jcount && !jh->b_transaction) {
1847                 __journal_remove_journal_head(bh);
1848                 __brelse(bh);
1849         }
1850         jbd_unlock_bh_journal_head(bh);
1851 }
1852
1853 /*
1854  * /proc tunables
1855  */
1856 #if defined(CONFIG_JBD_DEBUG)
1857 int journal_enable_debug;
1858 EXPORT_SYMBOL(journal_enable_debug);
1859 #endif
1860
1861 #if defined(CONFIG_JBD_DEBUG) && defined(CONFIG_PROC_FS)
1862
1863 static struct proc_dir_entry *proc_jbd_debug;
1864
1865 static int read_jbd_debug(char *page, char **start, off_t off,
1866                           int count, int *eof, void *data)
1867 {
1868         int ret;
1869
1870         ret = sprintf(page + off, "%d\n", journal_enable_debug);
1871         *eof = 1;
1872         return ret;
1873 }
1874
1875 static int write_jbd_debug(struct file *file, const char __user *buffer,
1876                            unsigned long count, void *data)
1877 {
1878         char buf[32];
1879
1880         if (count > ARRAY_SIZE(buf) - 1)
1881                 count = ARRAY_SIZE(buf) - 1;
1882         if (copy_from_user(buf, buffer, count))
1883                 return -EFAULT;
1884         buf[ARRAY_SIZE(buf) - 1] = '\0';
1885         journal_enable_debug = simple_strtoul(buf, NULL, 10);
1886         return count;
1887 }
1888
1889 #define JBD_PROC_NAME "sys/fs/jbd-debug"
1890
1891 static void __init create_jbd_proc_entry(void)
1892 {
1893         proc_jbd_debug = create_proc_entry(JBD_PROC_NAME, 0644, NULL);
1894         if (proc_jbd_debug) {
1895                 /* Why is this so hard? */
1896                 proc_jbd_debug->read_proc = read_jbd_debug;
1897                 proc_jbd_debug->write_proc = write_jbd_debug;
1898         }
1899 }
1900
1901 static void __exit remove_jbd_proc_entry(void)
1902 {
1903         if (proc_jbd_debug)
1904                 remove_proc_entry(JBD_PROC_NAME, NULL);
1905 }
1906
1907 #else
1908
1909 #define create_jbd_proc_entry() do {} while (0)
1910 #define remove_jbd_proc_entry() do {} while (0)
1911
1912 #endif
1913
1914 struct kmem_cache *jbd_handle_cache;
1915
1916 static int __init journal_init_handle_cache(void)
1917 {
1918         jbd_handle_cache = kmem_cache_create("journal_handle",
1919                                 sizeof(handle_t),
1920                                 0,              /* offset */
1921                                 SLAB_TEMPORARY, /* flags */
1922                                 NULL);          /* ctor */
1923         if (jbd_handle_cache == NULL) {
1924                 printk(KERN_EMERG "JBD: failed to create handle cache\n");
1925                 return -ENOMEM;
1926         }
1927         return 0;
1928 }
1929
1930 static void journal_destroy_handle_cache(void)
1931 {
1932         if (jbd_handle_cache)
1933                 kmem_cache_destroy(jbd_handle_cache);
1934 }
1935
1936 /*
1937  * Module startup and shutdown
1938  */
1939
1940 static int __init journal_init_caches(void)
1941 {
1942         int ret;
1943
1944         ret = journal_init_revoke_caches();
1945         if (ret == 0)
1946                 ret = journal_init_journal_head_cache();
1947         if (ret == 0)
1948                 ret = journal_init_handle_cache();
1949         return ret;
1950 }
1951
1952 static void journal_destroy_caches(void)
1953 {
1954         journal_destroy_revoke_caches();
1955         journal_destroy_journal_head_cache();
1956         journal_destroy_handle_cache();
1957 }
1958
1959 static int __init journal_init(void)
1960 {
1961         int ret;
1962
1963         BUILD_BUG_ON(sizeof(struct journal_superblock_s) != 1024);
1964
1965         ret = journal_init_caches();
1966         if (ret != 0)
1967                 journal_destroy_caches();
1968         create_jbd_proc_entry();
1969         return ret;
1970 }
1971
1972 static void __exit journal_exit(void)
1973 {
1974 #ifdef CONFIG_JBD_DEBUG
1975         int n = atomic_read(&nr_journal_heads);
1976         if (n)
1977                 printk(KERN_EMERG "JBD: leaked %d journal_heads!\n", n);
1978 #endif
1979         remove_jbd_proc_entry();
1980         journal_destroy_caches();
1981 }
1982
1983 MODULE_LICENSE("GPL");
1984 module_init(journal_init);
1985 module_exit(journal_exit);
1986