Merge branch 'for_linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tytso/ext4
[linux-2.6.git] / fs / jbd2 / journal.c
1 /*
2  * linux/fs/jbd2/journal.c
3  *
4  * Written by Stephen C. Tweedie <sct@redhat.com>, 1998
5  *
6  * Copyright 1998 Red Hat corp --- All Rights Reserved
7  *
8  * This file is part of the Linux kernel and is made available under
9  * the terms of the GNU General Public License, version 2, or at your
10  * option, any later version, incorporated herein by reference.
11  *
12  * Generic filesystem journal-writing code; part of the ext2fs
13  * journaling system.
14  *
15  * This file manages journals: areas of disk reserved for logging
16  * transactional updates.  This includes the kernel journaling thread
17  * which is responsible for scheduling updates to the log.
18  *
19  * We do not actually manage the physical storage of the journal in this
20  * file: that is left to a per-journal policy function, which allows us
21  * to store the journal within a filesystem-specified area for ext2
22  * journaling (ext2 can use a reserved inode for storing the log).
23  */
24
25 #include <linux/module.h>
26 #include <linux/time.h>
27 #include <linux/fs.h>
28 #include <linux/jbd2.h>
29 #include <linux/errno.h>
30 #include <linux/slab.h>
31 #include <linux/init.h>
32 #include <linux/mm.h>
33 #include <linux/freezer.h>
34 #include <linux/pagemap.h>
35 #include <linux/kthread.h>
36 #include <linux/poison.h>
37 #include <linux/proc_fs.h>
38 #include <linux/debugfs.h>
39 #include <linux/seq_file.h>
40 #include <linux/math64.h>
41 #include <linux/hash.h>
42 #include <linux/log2.h>
43 #include <linux/vmalloc.h>
44 #include <linux/backing-dev.h>
45 #include <linux/bitops.h>
46 #include <linux/ratelimit.h>
47
48 #define CREATE_TRACE_POINTS
49 #include <trace/events/jbd2.h>
50
51 #include <asm/uaccess.h>
52 #include <asm/page.h>
53 #include <asm/system.h>
54
55 EXPORT_SYMBOL(jbd2_journal_extend);
56 EXPORT_SYMBOL(jbd2_journal_stop);
57 EXPORT_SYMBOL(jbd2_journal_lock_updates);
58 EXPORT_SYMBOL(jbd2_journal_unlock_updates);
59 EXPORT_SYMBOL(jbd2_journal_get_write_access);
60 EXPORT_SYMBOL(jbd2_journal_get_create_access);
61 EXPORT_SYMBOL(jbd2_journal_get_undo_access);
62 EXPORT_SYMBOL(jbd2_journal_set_triggers);
63 EXPORT_SYMBOL(jbd2_journal_dirty_metadata);
64 EXPORT_SYMBOL(jbd2_journal_release_buffer);
65 EXPORT_SYMBOL(jbd2_journal_forget);
66 #if 0
67 EXPORT_SYMBOL(journal_sync_buffer);
68 #endif
69 EXPORT_SYMBOL(jbd2_journal_flush);
70 EXPORT_SYMBOL(jbd2_journal_revoke);
71
72 EXPORT_SYMBOL(jbd2_journal_init_dev);
73 EXPORT_SYMBOL(jbd2_journal_init_inode);
74 EXPORT_SYMBOL(jbd2_journal_update_format);
75 EXPORT_SYMBOL(jbd2_journal_check_used_features);
76 EXPORT_SYMBOL(jbd2_journal_check_available_features);
77 EXPORT_SYMBOL(jbd2_journal_set_features);
78 EXPORT_SYMBOL(jbd2_journal_load);
79 EXPORT_SYMBOL(jbd2_journal_destroy);
80 EXPORT_SYMBOL(jbd2_journal_abort);
81 EXPORT_SYMBOL(jbd2_journal_errno);
82 EXPORT_SYMBOL(jbd2_journal_ack_err);
83 EXPORT_SYMBOL(jbd2_journal_clear_err);
84 EXPORT_SYMBOL(jbd2_log_wait_commit);
85 EXPORT_SYMBOL(jbd2_log_start_commit);
86 EXPORT_SYMBOL(jbd2_journal_start_commit);
87 EXPORT_SYMBOL(jbd2_journal_force_commit_nested);
88 EXPORT_SYMBOL(jbd2_journal_wipe);
89 EXPORT_SYMBOL(jbd2_journal_blocks_per_page);
90 EXPORT_SYMBOL(jbd2_journal_invalidatepage);
91 EXPORT_SYMBOL(jbd2_journal_try_to_free_buffers);
92 EXPORT_SYMBOL(jbd2_journal_force_commit);
93 EXPORT_SYMBOL(jbd2_journal_file_inode);
94 EXPORT_SYMBOL(jbd2_journal_init_jbd_inode);
95 EXPORT_SYMBOL(jbd2_journal_release_jbd_inode);
96 EXPORT_SYMBOL(jbd2_journal_begin_ordered_truncate);
97 EXPORT_SYMBOL(jbd2_inode_cache);
98
99 static int journal_convert_superblock_v1(journal_t *, journal_superblock_t *);
100 static void __journal_abort_soft (journal_t *journal, int errno);
101 static int jbd2_journal_create_slab(size_t slab_size);
102
103 /*
104  * Helper function used to manage commit timeouts
105  */
106
107 static void commit_timeout(unsigned long __data)
108 {
109         struct task_struct * p = (struct task_struct *) __data;
110
111         wake_up_process(p);
112 }
113
114 /*
115  * kjournald2: The main thread function used to manage a logging device
116  * journal.
117  *
118  * This kernel thread is responsible for two things:
119  *
120  * 1) COMMIT:  Every so often we need to commit the current state of the
121  *    filesystem to disk.  The journal thread is responsible for writing
122  *    all of the metadata buffers to disk.
123  *
124  * 2) CHECKPOINT: We cannot reuse a used section of the log file until all
125  *    of the data in that part of the log has been rewritten elsewhere on
126  *    the disk.  Flushing these old buffers to reclaim space in the log is
127  *    known as checkpointing, and this thread is responsible for that job.
128  */
129
130 static int kjournald2(void *arg)
131 {
132         journal_t *journal = arg;
133         transaction_t *transaction;
134
135         /*
136          * Set up an interval timer which can be used to trigger a commit wakeup
137          * after the commit interval expires
138          */
139         setup_timer(&journal->j_commit_timer, commit_timeout,
140                         (unsigned long)current);
141
142         /* Record that the journal thread is running */
143         journal->j_task = current;
144         wake_up(&journal->j_wait_done_commit);
145
146         /*
147          * And now, wait forever for commit wakeup events.
148          */
149         write_lock(&journal->j_state_lock);
150
151 loop:
152         if (journal->j_flags & JBD2_UNMOUNT)
153                 goto end_loop;
154
155         jbd_debug(1, "commit_sequence=%d, commit_request=%d\n",
156                 journal->j_commit_sequence, journal->j_commit_request);
157
158         if (journal->j_commit_sequence != journal->j_commit_request) {
159                 jbd_debug(1, "OK, requests differ\n");
160                 write_unlock(&journal->j_state_lock);
161                 del_timer_sync(&journal->j_commit_timer);
162                 jbd2_journal_commit_transaction(journal);
163                 write_lock(&journal->j_state_lock);
164                 goto loop;
165         }
166
167         wake_up(&journal->j_wait_done_commit);
168         if (freezing(current)) {
169                 /*
170                  * The simpler the better. Flushing journal isn't a
171                  * good idea, because that depends on threads that may
172                  * be already stopped.
173                  */
174                 jbd_debug(1, "Now suspending kjournald2\n");
175                 write_unlock(&journal->j_state_lock);
176                 refrigerator();
177                 write_lock(&journal->j_state_lock);
178         } else {
179                 /*
180                  * We assume on resume that commits are already there,
181                  * so we don't sleep
182                  */
183                 DEFINE_WAIT(wait);
184                 int should_sleep = 1;
185
186                 prepare_to_wait(&journal->j_wait_commit, &wait,
187                                 TASK_INTERRUPTIBLE);
188                 if (journal->j_commit_sequence != journal->j_commit_request)
189                         should_sleep = 0;
190                 transaction = journal->j_running_transaction;
191                 if (transaction && time_after_eq(jiffies,
192                                                 transaction->t_expires))
193                         should_sleep = 0;
194                 if (journal->j_flags & JBD2_UNMOUNT)
195                         should_sleep = 0;
196                 if (should_sleep) {
197                         write_unlock(&journal->j_state_lock);
198                         schedule();
199                         write_lock(&journal->j_state_lock);
200                 }
201                 finish_wait(&journal->j_wait_commit, &wait);
202         }
203
204         jbd_debug(1, "kjournald2 wakes\n");
205
206         /*
207          * Were we woken up by a commit wakeup event?
208          */
209         transaction = journal->j_running_transaction;
210         if (transaction && time_after_eq(jiffies, transaction->t_expires)) {
211                 journal->j_commit_request = transaction->t_tid;
212                 jbd_debug(1, "woke because of timeout\n");
213         }
214         goto loop;
215
216 end_loop:
217         write_unlock(&journal->j_state_lock);
218         del_timer_sync(&journal->j_commit_timer);
219         journal->j_task = NULL;
220         wake_up(&journal->j_wait_done_commit);
221         jbd_debug(1, "Journal thread exiting.\n");
222         return 0;
223 }
224
225 static int jbd2_journal_start_thread(journal_t *journal)
226 {
227         struct task_struct *t;
228
229         t = kthread_run(kjournald2, journal, "jbd2/%s",
230                         journal->j_devname);
231         if (IS_ERR(t))
232                 return PTR_ERR(t);
233
234         wait_event(journal->j_wait_done_commit, journal->j_task != NULL);
235         return 0;
236 }
237
238 static void journal_kill_thread(journal_t *journal)
239 {
240         write_lock(&journal->j_state_lock);
241         journal->j_flags |= JBD2_UNMOUNT;
242
243         while (journal->j_task) {
244                 wake_up(&journal->j_wait_commit);
245                 write_unlock(&journal->j_state_lock);
246                 wait_event(journal->j_wait_done_commit, journal->j_task == NULL);
247                 write_lock(&journal->j_state_lock);
248         }
249         write_unlock(&journal->j_state_lock);
250 }
251
252 /*
253  * jbd2_journal_write_metadata_buffer: write a metadata buffer to the journal.
254  *
255  * Writes a metadata buffer to a given disk block.  The actual IO is not
256  * performed but a new buffer_head is constructed which labels the data
257  * to be written with the correct destination disk block.
258  *
259  * Any magic-number escaping which needs to be done will cause a
260  * copy-out here.  If the buffer happens to start with the
261  * JBD2_MAGIC_NUMBER, then we can't write it to the log directly: the
262  * magic number is only written to the log for descripter blocks.  In
263  * this case, we copy the data and replace the first word with 0, and we
264  * return a result code which indicates that this buffer needs to be
265  * marked as an escaped buffer in the corresponding log descriptor
266  * block.  The missing word can then be restored when the block is read
267  * during recovery.
268  *
269  * If the source buffer has already been modified by a new transaction
270  * since we took the last commit snapshot, we use the frozen copy of
271  * that data for IO.  If we end up using the existing buffer_head's data
272  * for the write, then we *have* to lock the buffer to prevent anyone
273  * else from using and possibly modifying it while the IO is in
274  * progress.
275  *
276  * The function returns a pointer to the buffer_heads to be used for IO.
277  *
278  * We assume that the journal has already been locked in this function.
279  *
280  * Return value:
281  *  <0: Error
282  * >=0: Finished OK
283  *
284  * On success:
285  * Bit 0 set == escape performed on the data
286  * Bit 1 set == buffer copy-out performed (kfree the data after IO)
287  */
288
289 int jbd2_journal_write_metadata_buffer(transaction_t *transaction,
290                                   struct journal_head  *jh_in,
291                                   struct journal_head **jh_out,
292                                   unsigned long long blocknr)
293 {
294         int need_copy_out = 0;
295         int done_copy_out = 0;
296         int do_escape = 0;
297         char *mapped_data;
298         struct buffer_head *new_bh;
299         struct journal_head *new_jh;
300         struct page *new_page;
301         unsigned int new_offset;
302         struct buffer_head *bh_in = jh2bh(jh_in);
303         journal_t *journal = transaction->t_journal;
304
305         /*
306          * The buffer really shouldn't be locked: only the current committing
307          * transaction is allowed to write it, so nobody else is allowed
308          * to do any IO.
309          *
310          * akpm: except if we're journalling data, and write() output is
311          * also part of a shared mapping, and another thread has
312          * decided to launch a writepage() against this buffer.
313          */
314         J_ASSERT_BH(bh_in, buffer_jbddirty(bh_in));
315
316 retry_alloc:
317         new_bh = alloc_buffer_head(GFP_NOFS);
318         if (!new_bh) {
319                 /*
320                  * Failure is not an option, but __GFP_NOFAIL is going
321                  * away; so we retry ourselves here.
322                  */
323                 congestion_wait(BLK_RW_ASYNC, HZ/50);
324                 goto retry_alloc;
325         }
326
327         /* keep subsequent assertions sane */
328         new_bh->b_state = 0;
329         init_buffer(new_bh, NULL, NULL);
330         atomic_set(&new_bh->b_count, 1);
331         new_jh = jbd2_journal_add_journal_head(new_bh); /* This sleeps */
332
333         /*
334          * If a new transaction has already done a buffer copy-out, then
335          * we use that version of the data for the commit.
336          */
337         jbd_lock_bh_state(bh_in);
338 repeat:
339         if (jh_in->b_frozen_data) {
340                 done_copy_out = 1;
341                 new_page = virt_to_page(jh_in->b_frozen_data);
342                 new_offset = offset_in_page(jh_in->b_frozen_data);
343         } else {
344                 new_page = jh2bh(jh_in)->b_page;
345                 new_offset = offset_in_page(jh2bh(jh_in)->b_data);
346         }
347
348         mapped_data = kmap_atomic(new_page, KM_USER0);
349         /*
350          * Fire data frozen trigger if data already wasn't frozen.  Do this
351          * before checking for escaping, as the trigger may modify the magic
352          * offset.  If a copy-out happens afterwards, it will have the correct
353          * data in the buffer.
354          */
355         if (!done_copy_out)
356                 jbd2_buffer_frozen_trigger(jh_in, mapped_data + new_offset,
357                                            jh_in->b_triggers);
358
359         /*
360          * Check for escaping
361          */
362         if (*((__be32 *)(mapped_data + new_offset)) ==
363                                 cpu_to_be32(JBD2_MAGIC_NUMBER)) {
364                 need_copy_out = 1;
365                 do_escape = 1;
366         }
367         kunmap_atomic(mapped_data, KM_USER0);
368
369         /*
370          * Do we need to do a data copy?
371          */
372         if (need_copy_out && !done_copy_out) {
373                 char *tmp;
374
375                 jbd_unlock_bh_state(bh_in);
376                 tmp = jbd2_alloc(bh_in->b_size, GFP_NOFS);
377                 if (!tmp) {
378                         jbd2_journal_put_journal_head(new_jh);
379                         return -ENOMEM;
380                 }
381                 jbd_lock_bh_state(bh_in);
382                 if (jh_in->b_frozen_data) {
383                         jbd2_free(tmp, bh_in->b_size);
384                         goto repeat;
385                 }
386
387                 jh_in->b_frozen_data = tmp;
388                 mapped_data = kmap_atomic(new_page, KM_USER0);
389                 memcpy(tmp, mapped_data + new_offset, jh2bh(jh_in)->b_size);
390                 kunmap_atomic(mapped_data, KM_USER0);
391
392                 new_page = virt_to_page(tmp);
393                 new_offset = offset_in_page(tmp);
394                 done_copy_out = 1;
395
396                 /*
397                  * This isn't strictly necessary, as we're using frozen
398                  * data for the escaping, but it keeps consistency with
399                  * b_frozen_data usage.
400                  */
401                 jh_in->b_frozen_triggers = jh_in->b_triggers;
402         }
403
404         /*
405          * Did we need to do an escaping?  Now we've done all the
406          * copying, we can finally do so.
407          */
408         if (do_escape) {
409                 mapped_data = kmap_atomic(new_page, KM_USER0);
410                 *((unsigned int *)(mapped_data + new_offset)) = 0;
411                 kunmap_atomic(mapped_data, KM_USER0);
412         }
413
414         set_bh_page(new_bh, new_page, new_offset);
415         new_jh->b_transaction = NULL;
416         new_bh->b_size = jh2bh(jh_in)->b_size;
417         new_bh->b_bdev = transaction->t_journal->j_dev;
418         new_bh->b_blocknr = blocknr;
419         set_buffer_mapped(new_bh);
420         set_buffer_dirty(new_bh);
421
422         *jh_out = new_jh;
423
424         /*
425          * The to-be-written buffer needs to get moved to the io queue,
426          * and the original buffer whose contents we are shadowing or
427          * copying is moved to the transaction's shadow queue.
428          */
429         JBUFFER_TRACE(jh_in, "file as BJ_Shadow");
430         spin_lock(&journal->j_list_lock);
431         __jbd2_journal_file_buffer(jh_in, transaction, BJ_Shadow);
432         spin_unlock(&journal->j_list_lock);
433         jbd_unlock_bh_state(bh_in);
434
435         JBUFFER_TRACE(new_jh, "file as BJ_IO");
436         jbd2_journal_file_buffer(new_jh, transaction, BJ_IO);
437
438         return do_escape | (done_copy_out << 1);
439 }
440
441 /*
442  * Allocation code for the journal file.  Manage the space left in the
443  * journal, so that we can begin checkpointing when appropriate.
444  */
445
446 /*
447  * __jbd2_log_space_left: Return the number of free blocks left in the journal.
448  *
449  * Called with the journal already locked.
450  *
451  * Called under j_state_lock
452  */
453
454 int __jbd2_log_space_left(journal_t *journal)
455 {
456         int left = journal->j_free;
457
458         /* assert_spin_locked(&journal->j_state_lock); */
459
460         /*
461          * Be pessimistic here about the number of those free blocks which
462          * might be required for log descriptor control blocks.
463          */
464
465 #define MIN_LOG_RESERVED_BLOCKS 32 /* Allow for rounding errors */
466
467         left -= MIN_LOG_RESERVED_BLOCKS;
468
469         if (left <= 0)
470                 return 0;
471         left -= (left >> 3);
472         return left;
473 }
474
475 /*
476  * Called with j_state_lock locked for writing.
477  * Returns true if a transaction commit was started.
478  */
479 int __jbd2_log_start_commit(journal_t *journal, tid_t target)
480 {
481         /*
482          * Are we already doing a recent enough commit?
483          */
484         if (!tid_geq(journal->j_commit_request, target)) {
485                 /*
486                  * We want a new commit: OK, mark the request and wakeup the
487                  * commit thread.  We do _not_ do the commit ourselves.
488                  */
489
490                 journal->j_commit_request = target;
491                 jbd_debug(1, "JBD: requesting commit %d/%d\n",
492                           journal->j_commit_request,
493                           journal->j_commit_sequence);
494                 wake_up(&journal->j_wait_commit);
495                 return 1;
496         }
497         return 0;
498 }
499
500 int jbd2_log_start_commit(journal_t *journal, tid_t tid)
501 {
502         int ret;
503
504         write_lock(&journal->j_state_lock);
505         ret = __jbd2_log_start_commit(journal, tid);
506         write_unlock(&journal->j_state_lock);
507         return ret;
508 }
509
510 /*
511  * Force and wait upon a commit if the calling process is not within
512  * transaction.  This is used for forcing out undo-protected data which contains
513  * bitmaps, when the fs is running out of space.
514  *
515  * We can only force the running transaction if we don't have an active handle;
516  * otherwise, we will deadlock.
517  *
518  * Returns true if a transaction was started.
519  */
520 int jbd2_journal_force_commit_nested(journal_t *journal)
521 {
522         transaction_t *transaction = NULL;
523         tid_t tid;
524         int need_to_start = 0;
525
526         read_lock(&journal->j_state_lock);
527         if (journal->j_running_transaction && !current->journal_info) {
528                 transaction = journal->j_running_transaction;
529                 if (!tid_geq(journal->j_commit_request, transaction->t_tid))
530                         need_to_start = 1;
531         } else if (journal->j_committing_transaction)
532                 transaction = journal->j_committing_transaction;
533
534         if (!transaction) {
535                 read_unlock(&journal->j_state_lock);
536                 return 0;       /* Nothing to retry */
537         }
538
539         tid = transaction->t_tid;
540         read_unlock(&journal->j_state_lock);
541         if (need_to_start)
542                 jbd2_log_start_commit(journal, tid);
543         jbd2_log_wait_commit(journal, tid);
544         return 1;
545 }
546
547 /*
548  * Start a commit of the current running transaction (if any).  Returns true
549  * if a transaction is going to be committed (or is currently already
550  * committing), and fills its tid in at *ptid
551  */
552 int jbd2_journal_start_commit(journal_t *journal, tid_t *ptid)
553 {
554         int ret = 0;
555
556         write_lock(&journal->j_state_lock);
557         if (journal->j_running_transaction) {
558                 tid_t tid = journal->j_running_transaction->t_tid;
559
560                 __jbd2_log_start_commit(journal, tid);
561                 /* There's a running transaction and we've just made sure
562                  * it's commit has been scheduled. */
563                 if (ptid)
564                         *ptid = tid;
565                 ret = 1;
566         } else if (journal->j_committing_transaction) {
567                 /*
568                  * If ext3_write_super() recently started a commit, then we
569                  * have to wait for completion of that transaction
570                  */
571                 if (ptid)
572                         *ptid = journal->j_committing_transaction->t_tid;
573                 ret = 1;
574         }
575         write_unlock(&journal->j_state_lock);
576         return ret;
577 }
578
579 /*
580  * Wait for a specified commit to complete.
581  * The caller may not hold the journal lock.
582  */
583 int jbd2_log_wait_commit(journal_t *journal, tid_t tid)
584 {
585         int err = 0;
586
587         read_lock(&journal->j_state_lock);
588 #ifdef CONFIG_JBD2_DEBUG
589         if (!tid_geq(journal->j_commit_request, tid)) {
590                 printk(KERN_EMERG
591                        "%s: error: j_commit_request=%d, tid=%d\n",
592                        __func__, journal->j_commit_request, tid);
593         }
594 #endif
595         while (tid_gt(tid, journal->j_commit_sequence)) {
596                 jbd_debug(1, "JBD: want %d, j_commit_sequence=%d\n",
597                                   tid, journal->j_commit_sequence);
598                 wake_up(&journal->j_wait_commit);
599                 read_unlock(&journal->j_state_lock);
600                 wait_event(journal->j_wait_done_commit,
601                                 !tid_gt(tid, journal->j_commit_sequence));
602                 read_lock(&journal->j_state_lock);
603         }
604         read_unlock(&journal->j_state_lock);
605
606         if (unlikely(is_journal_aborted(journal))) {
607                 printk(KERN_EMERG "journal commit I/O error\n");
608                 err = -EIO;
609         }
610         return err;
611 }
612
613 /*
614  * Log buffer allocation routines:
615  */
616
617 int jbd2_journal_next_log_block(journal_t *journal, unsigned long long *retp)
618 {
619         unsigned long blocknr;
620
621         write_lock(&journal->j_state_lock);
622         J_ASSERT(journal->j_free > 1);
623
624         blocknr = journal->j_head;
625         journal->j_head++;
626         journal->j_free--;
627         if (journal->j_head == journal->j_last)
628                 journal->j_head = journal->j_first;
629         write_unlock(&journal->j_state_lock);
630         return jbd2_journal_bmap(journal, blocknr, retp);
631 }
632
633 /*
634  * Conversion of logical to physical block numbers for the journal
635  *
636  * On external journals the journal blocks are identity-mapped, so
637  * this is a no-op.  If needed, we can use j_blk_offset - everything is
638  * ready.
639  */
640 int jbd2_journal_bmap(journal_t *journal, unsigned long blocknr,
641                  unsigned long long *retp)
642 {
643         int err = 0;
644         unsigned long long ret;
645
646         if (journal->j_inode) {
647                 ret = bmap(journal->j_inode, blocknr);
648                 if (ret)
649                         *retp = ret;
650                 else {
651                         printk(KERN_ALERT "%s: journal block not found "
652                                         "at offset %lu on %s\n",
653                                __func__, blocknr, journal->j_devname);
654                         err = -EIO;
655                         __journal_abort_soft(journal, err);
656                 }
657         } else {
658                 *retp = blocknr; /* +journal->j_blk_offset */
659         }
660         return err;
661 }
662
663 /*
664  * We play buffer_head aliasing tricks to write data/metadata blocks to
665  * the journal without copying their contents, but for journal
666  * descriptor blocks we do need to generate bona fide buffers.
667  *
668  * After the caller of jbd2_journal_get_descriptor_buffer() has finished modifying
669  * the buffer's contents they really should run flush_dcache_page(bh->b_page).
670  * But we don't bother doing that, so there will be coherency problems with
671  * mmaps of blockdevs which hold live JBD-controlled filesystems.
672  */
673 struct journal_head *jbd2_journal_get_descriptor_buffer(journal_t *journal)
674 {
675         struct buffer_head *bh;
676         unsigned long long blocknr;
677         int err;
678
679         err = jbd2_journal_next_log_block(journal, &blocknr);
680
681         if (err)
682                 return NULL;
683
684         bh = __getblk(journal->j_dev, blocknr, journal->j_blocksize);
685         if (!bh)
686                 return NULL;
687         lock_buffer(bh);
688         memset(bh->b_data, 0, journal->j_blocksize);
689         set_buffer_uptodate(bh);
690         unlock_buffer(bh);
691         BUFFER_TRACE(bh, "return this buffer");
692         return jbd2_journal_add_journal_head(bh);
693 }
694
695 struct jbd2_stats_proc_session {
696         journal_t *journal;
697         struct transaction_stats_s *stats;
698         int start;
699         int max;
700 };
701
702 static void *jbd2_seq_info_start(struct seq_file *seq, loff_t *pos)
703 {
704         return *pos ? NULL : SEQ_START_TOKEN;
705 }
706
707 static void *jbd2_seq_info_next(struct seq_file *seq, void *v, loff_t *pos)
708 {
709         return NULL;
710 }
711
712 static int jbd2_seq_info_show(struct seq_file *seq, void *v)
713 {
714         struct jbd2_stats_proc_session *s = seq->private;
715
716         if (v != SEQ_START_TOKEN)
717                 return 0;
718         seq_printf(seq, "%lu transaction, each up to %u blocks\n",
719                         s->stats->ts_tid,
720                         s->journal->j_max_transaction_buffers);
721         if (s->stats->ts_tid == 0)
722                 return 0;
723         seq_printf(seq, "average: \n  %ums waiting for transaction\n",
724             jiffies_to_msecs(s->stats->run.rs_wait / s->stats->ts_tid));
725         seq_printf(seq, "  %ums running transaction\n",
726             jiffies_to_msecs(s->stats->run.rs_running / s->stats->ts_tid));
727         seq_printf(seq, "  %ums transaction was being locked\n",
728             jiffies_to_msecs(s->stats->run.rs_locked / s->stats->ts_tid));
729         seq_printf(seq, "  %ums flushing data (in ordered mode)\n",
730             jiffies_to_msecs(s->stats->run.rs_flushing / s->stats->ts_tid));
731         seq_printf(seq, "  %ums logging transaction\n",
732             jiffies_to_msecs(s->stats->run.rs_logging / s->stats->ts_tid));
733         seq_printf(seq, "  %lluus average transaction commit time\n",
734                    div_u64(s->journal->j_average_commit_time, 1000));
735         seq_printf(seq, "  %lu handles per transaction\n",
736             s->stats->run.rs_handle_count / s->stats->ts_tid);
737         seq_printf(seq, "  %lu blocks per transaction\n",
738             s->stats->run.rs_blocks / s->stats->ts_tid);
739         seq_printf(seq, "  %lu logged blocks per transaction\n",
740             s->stats->run.rs_blocks_logged / s->stats->ts_tid);
741         return 0;
742 }
743
744 static void jbd2_seq_info_stop(struct seq_file *seq, void *v)
745 {
746 }
747
748 static const struct seq_operations jbd2_seq_info_ops = {
749         .start  = jbd2_seq_info_start,
750         .next   = jbd2_seq_info_next,
751         .stop   = jbd2_seq_info_stop,
752         .show   = jbd2_seq_info_show,
753 };
754
755 static int jbd2_seq_info_open(struct inode *inode, struct file *file)
756 {
757         journal_t *journal = PDE(inode)->data;
758         struct jbd2_stats_proc_session *s;
759         int rc, size;
760
761         s = kmalloc(sizeof(*s), GFP_KERNEL);
762         if (s == NULL)
763                 return -ENOMEM;
764         size = sizeof(struct transaction_stats_s);
765         s->stats = kmalloc(size, GFP_KERNEL);
766         if (s->stats == NULL) {
767                 kfree(s);
768                 return -ENOMEM;
769         }
770         spin_lock(&journal->j_history_lock);
771         memcpy(s->stats, &journal->j_stats, size);
772         s->journal = journal;
773         spin_unlock(&journal->j_history_lock);
774
775         rc = seq_open(file, &jbd2_seq_info_ops);
776         if (rc == 0) {
777                 struct seq_file *m = file->private_data;
778                 m->private = s;
779         } else {
780                 kfree(s->stats);
781                 kfree(s);
782         }
783         return rc;
784
785 }
786
787 static int jbd2_seq_info_release(struct inode *inode, struct file *file)
788 {
789         struct seq_file *seq = file->private_data;
790         struct jbd2_stats_proc_session *s = seq->private;
791         kfree(s->stats);
792         kfree(s);
793         return seq_release(inode, file);
794 }
795
796 static const struct file_operations jbd2_seq_info_fops = {
797         .owner          = THIS_MODULE,
798         .open           = jbd2_seq_info_open,
799         .read           = seq_read,
800         .llseek         = seq_lseek,
801         .release        = jbd2_seq_info_release,
802 };
803
804 static struct proc_dir_entry *proc_jbd2_stats;
805
806 static void jbd2_stats_proc_init(journal_t *journal)
807 {
808         journal->j_proc_entry = proc_mkdir(journal->j_devname, proc_jbd2_stats);
809         if (journal->j_proc_entry) {
810                 proc_create_data("info", S_IRUGO, journal->j_proc_entry,
811                                  &jbd2_seq_info_fops, journal);
812         }
813 }
814
815 static void jbd2_stats_proc_exit(journal_t *journal)
816 {
817         remove_proc_entry("info", journal->j_proc_entry);
818         remove_proc_entry(journal->j_devname, proc_jbd2_stats);
819 }
820
821 /*
822  * Management for journal control blocks: functions to create and
823  * destroy journal_t structures, and to initialise and read existing
824  * journal blocks from disk.  */
825
826 /* First: create and setup a journal_t object in memory.  We initialise
827  * very few fields yet: that has to wait until we have created the
828  * journal structures from from scratch, or loaded them from disk. */
829
830 static journal_t * journal_init_common (void)
831 {
832         journal_t *journal;
833         int err;
834
835         journal = kzalloc(sizeof(*journal), GFP_KERNEL);
836         if (!journal)
837                 return NULL;
838
839         init_waitqueue_head(&journal->j_wait_transaction_locked);
840         init_waitqueue_head(&journal->j_wait_logspace);
841         init_waitqueue_head(&journal->j_wait_done_commit);
842         init_waitqueue_head(&journal->j_wait_checkpoint);
843         init_waitqueue_head(&journal->j_wait_commit);
844         init_waitqueue_head(&journal->j_wait_updates);
845         mutex_init(&journal->j_barrier);
846         mutex_init(&journal->j_checkpoint_mutex);
847         spin_lock_init(&journal->j_revoke_lock);
848         spin_lock_init(&journal->j_list_lock);
849         rwlock_init(&journal->j_state_lock);
850
851         journal->j_commit_interval = (HZ * JBD2_DEFAULT_MAX_COMMIT_AGE);
852         journal->j_min_batch_time = 0;
853         journal->j_max_batch_time = 15000; /* 15ms */
854
855         /* The journal is marked for error until we succeed with recovery! */
856         journal->j_flags = JBD2_ABORT;
857
858         /* Set up a default-sized revoke table for the new mount. */
859         err = jbd2_journal_init_revoke(journal, JOURNAL_REVOKE_DEFAULT_HASH);
860         if (err) {
861                 kfree(journal);
862                 return NULL;
863         }
864
865         spin_lock_init(&journal->j_history_lock);
866
867         return journal;
868 }
869
870 /* jbd2_journal_init_dev and jbd2_journal_init_inode:
871  *
872  * Create a journal structure assigned some fixed set of disk blocks to
873  * the journal.  We don't actually touch those disk blocks yet, but we
874  * need to set up all of the mapping information to tell the journaling
875  * system where the journal blocks are.
876  *
877  */
878
879 /**
880  *  journal_t * jbd2_journal_init_dev() - creates and initialises a journal structure
881  *  @bdev: Block device on which to create the journal
882  *  @fs_dev: Device which hold journalled filesystem for this journal.
883  *  @start: Block nr Start of journal.
884  *  @len:  Length of the journal in blocks.
885  *  @blocksize: blocksize of journalling device
886  *
887  *  Returns: a newly created journal_t *
888  *
889  *  jbd2_journal_init_dev creates a journal which maps a fixed contiguous
890  *  range of blocks on an arbitrary block device.
891  *
892  */
893 journal_t * jbd2_journal_init_dev(struct block_device *bdev,
894                         struct block_device *fs_dev,
895                         unsigned long long start, int len, int blocksize)
896 {
897         journal_t *journal = journal_init_common();
898         struct buffer_head *bh;
899         char *p;
900         int n;
901
902         if (!journal)
903                 return NULL;
904
905         /* journal descriptor can store up to n blocks -bzzz */
906         journal->j_blocksize = blocksize;
907         journal->j_dev = bdev;
908         journal->j_fs_dev = fs_dev;
909         journal->j_blk_offset = start;
910         journal->j_maxlen = len;
911         bdevname(journal->j_dev, journal->j_devname);
912         p = journal->j_devname;
913         while ((p = strchr(p, '/')))
914                 *p = '!';
915         jbd2_stats_proc_init(journal);
916         n = journal->j_blocksize / sizeof(journal_block_tag_t);
917         journal->j_wbufsize = n;
918         journal->j_wbuf = kmalloc(n * sizeof(struct buffer_head*), GFP_KERNEL);
919         if (!journal->j_wbuf) {
920                 printk(KERN_ERR "%s: Can't allocate bhs for commit thread\n",
921                         __func__);
922                 goto out_err;
923         }
924
925         bh = __getblk(journal->j_dev, start, journal->j_blocksize);
926         if (!bh) {
927                 printk(KERN_ERR
928                        "%s: Cannot get buffer for journal superblock\n",
929                        __func__);
930                 goto out_err;
931         }
932         journal->j_sb_buffer = bh;
933         journal->j_superblock = (journal_superblock_t *)bh->b_data;
934
935         return journal;
936 out_err:
937         kfree(journal->j_wbuf);
938         jbd2_stats_proc_exit(journal);
939         kfree(journal);
940         return NULL;
941 }
942
943 /**
944  *  journal_t * jbd2_journal_init_inode () - creates a journal which maps to a inode.
945  *  @inode: An inode to create the journal in
946  *
947  * jbd2_journal_init_inode creates a journal which maps an on-disk inode as
948  * the journal.  The inode must exist already, must support bmap() and
949  * must have all data blocks preallocated.
950  */
951 journal_t * jbd2_journal_init_inode (struct inode *inode)
952 {
953         struct buffer_head *bh;
954         journal_t *journal = journal_init_common();
955         char *p;
956         int err;
957         int n;
958         unsigned long long blocknr;
959
960         if (!journal)
961                 return NULL;
962
963         journal->j_dev = journal->j_fs_dev = inode->i_sb->s_bdev;
964         journal->j_inode = inode;
965         bdevname(journal->j_dev, journal->j_devname);
966         p = journal->j_devname;
967         while ((p = strchr(p, '/')))
968                 *p = '!';
969         p = journal->j_devname + strlen(journal->j_devname);
970         sprintf(p, "-%lu", journal->j_inode->i_ino);
971         jbd_debug(1,
972                   "journal %p: inode %s/%ld, size %Ld, bits %d, blksize %ld\n",
973                   journal, inode->i_sb->s_id, inode->i_ino,
974                   (long long) inode->i_size,
975                   inode->i_sb->s_blocksize_bits, inode->i_sb->s_blocksize);
976
977         journal->j_maxlen = inode->i_size >> inode->i_sb->s_blocksize_bits;
978         journal->j_blocksize = inode->i_sb->s_blocksize;
979         jbd2_stats_proc_init(journal);
980
981         /* journal descriptor can store up to n blocks -bzzz */
982         n = journal->j_blocksize / sizeof(journal_block_tag_t);
983         journal->j_wbufsize = n;
984         journal->j_wbuf = kmalloc(n * sizeof(struct buffer_head*), GFP_KERNEL);
985         if (!journal->j_wbuf) {
986                 printk(KERN_ERR "%s: Can't allocate bhs for commit thread\n",
987                         __func__);
988                 goto out_err;
989         }
990
991         err = jbd2_journal_bmap(journal, 0, &blocknr);
992         /* If that failed, give up */
993         if (err) {
994                 printk(KERN_ERR "%s: Cannot locate journal superblock\n",
995                        __func__);
996                 goto out_err;
997         }
998
999         bh = __getblk(journal->j_dev, blocknr, journal->j_blocksize);
1000         if (!bh) {
1001                 printk(KERN_ERR
1002                        "%s: Cannot get buffer for journal superblock\n",
1003                        __func__);
1004                 goto out_err;
1005         }
1006         journal->j_sb_buffer = bh;
1007         journal->j_superblock = (journal_superblock_t *)bh->b_data;
1008
1009         return journal;
1010 out_err:
1011         kfree(journal->j_wbuf);
1012         jbd2_stats_proc_exit(journal);
1013         kfree(journal);
1014         return NULL;
1015 }
1016
1017 /*
1018  * If the journal init or create aborts, we need to mark the journal
1019  * superblock as being NULL to prevent the journal destroy from writing
1020  * back a bogus superblock.
1021  */
1022 static void journal_fail_superblock (journal_t *journal)
1023 {
1024         struct buffer_head *bh = journal->j_sb_buffer;
1025         brelse(bh);
1026         journal->j_sb_buffer = NULL;
1027 }
1028
1029 /*
1030  * Given a journal_t structure, initialise the various fields for
1031  * startup of a new journaling session.  We use this both when creating
1032  * a journal, and after recovering an old journal to reset it for
1033  * subsequent use.
1034  */
1035
1036 static int journal_reset(journal_t *journal)
1037 {
1038         journal_superblock_t *sb = journal->j_superblock;
1039         unsigned long long first, last;
1040
1041         first = be32_to_cpu(sb->s_first);
1042         last = be32_to_cpu(sb->s_maxlen);
1043         if (first + JBD2_MIN_JOURNAL_BLOCKS > last + 1) {
1044                 printk(KERN_ERR "JBD: Journal too short (blocks %llu-%llu).\n",
1045                        first, last);
1046                 journal_fail_superblock(journal);
1047                 return -EINVAL;
1048         }
1049
1050         journal->j_first = first;
1051         journal->j_last = last;
1052
1053         journal->j_head = first;
1054         journal->j_tail = first;
1055         journal->j_free = last - first;
1056
1057         journal->j_tail_sequence = journal->j_transaction_sequence;
1058         journal->j_commit_sequence = journal->j_transaction_sequence - 1;
1059         journal->j_commit_request = journal->j_commit_sequence;
1060
1061         journal->j_max_transaction_buffers = journal->j_maxlen / 4;
1062
1063         /* Add the dynamic fields and write it to disk. */
1064         jbd2_journal_update_superblock(journal, 1);
1065         return jbd2_journal_start_thread(journal);
1066 }
1067
1068 /**
1069  * void jbd2_journal_update_superblock() - Update journal sb on disk.
1070  * @journal: The journal to update.
1071  * @wait: Set to '0' if you don't want to wait for IO completion.
1072  *
1073  * Update a journal's dynamic superblock fields and write it to disk,
1074  * optionally waiting for the IO to complete.
1075  */
1076 void jbd2_journal_update_superblock(journal_t *journal, int wait)
1077 {
1078         journal_superblock_t *sb = journal->j_superblock;
1079         struct buffer_head *bh = journal->j_sb_buffer;
1080
1081         /*
1082          * As a special case, if the on-disk copy is already marked as needing
1083          * no recovery (s_start == 0) and there are no outstanding transactions
1084          * in the filesystem, then we can safely defer the superblock update
1085          * until the next commit by setting JBD2_FLUSHED.  This avoids
1086          * attempting a write to a potential-readonly device.
1087          */
1088         if (sb->s_start == 0 && journal->j_tail_sequence ==
1089                                 journal->j_transaction_sequence) {
1090                 jbd_debug(1,"JBD: Skipping superblock update on recovered sb "
1091                         "(start %ld, seq %d, errno %d)\n",
1092                         journal->j_tail, journal->j_tail_sequence,
1093                         journal->j_errno);
1094                 goto out;
1095         }
1096
1097         if (buffer_write_io_error(bh)) {
1098                 /*
1099                  * Oh, dear.  A previous attempt to write the journal
1100                  * superblock failed.  This could happen because the
1101                  * USB device was yanked out.  Or it could happen to
1102                  * be a transient write error and maybe the block will
1103                  * be remapped.  Nothing we can do but to retry the
1104                  * write and hope for the best.
1105                  */
1106                 printk(KERN_ERR "JBD2: previous I/O error detected "
1107                        "for journal superblock update for %s.\n",
1108                        journal->j_devname);
1109                 clear_buffer_write_io_error(bh);
1110                 set_buffer_uptodate(bh);
1111         }
1112
1113         read_lock(&journal->j_state_lock);
1114         jbd_debug(1,"JBD: updating superblock (start %ld, seq %d, errno %d)\n",
1115                   journal->j_tail, journal->j_tail_sequence, journal->j_errno);
1116
1117         sb->s_sequence = cpu_to_be32(journal->j_tail_sequence);
1118         sb->s_start    = cpu_to_be32(journal->j_tail);
1119         sb->s_errno    = cpu_to_be32(journal->j_errno);
1120         read_unlock(&journal->j_state_lock);
1121
1122         BUFFER_TRACE(bh, "marking dirty");
1123         mark_buffer_dirty(bh);
1124         if (wait) {
1125                 sync_dirty_buffer(bh);
1126                 if (buffer_write_io_error(bh)) {
1127                         printk(KERN_ERR "JBD2: I/O error detected "
1128                                "when updating journal superblock for %s.\n",
1129                                journal->j_devname);
1130                         clear_buffer_write_io_error(bh);
1131                         set_buffer_uptodate(bh);
1132                 }
1133         } else
1134                 write_dirty_buffer(bh, WRITE);
1135
1136 out:
1137         /* If we have just flushed the log (by marking s_start==0), then
1138          * any future commit will have to be careful to update the
1139          * superblock again to re-record the true start of the log. */
1140
1141         write_lock(&journal->j_state_lock);
1142         if (sb->s_start)
1143                 journal->j_flags &= ~JBD2_FLUSHED;
1144         else
1145                 journal->j_flags |= JBD2_FLUSHED;
1146         write_unlock(&journal->j_state_lock);
1147 }
1148
1149 /*
1150  * Read the superblock for a given journal, performing initial
1151  * validation of the format.
1152  */
1153
1154 static int journal_get_superblock(journal_t *journal)
1155 {
1156         struct buffer_head *bh;
1157         journal_superblock_t *sb;
1158         int err = -EIO;
1159
1160         bh = journal->j_sb_buffer;
1161
1162         J_ASSERT(bh != NULL);
1163         if (!buffer_uptodate(bh)) {
1164                 ll_rw_block(READ, 1, &bh);
1165                 wait_on_buffer(bh);
1166                 if (!buffer_uptodate(bh)) {
1167                         printk (KERN_ERR
1168                                 "JBD: IO error reading journal superblock\n");
1169                         goto out;
1170                 }
1171         }
1172
1173         sb = journal->j_superblock;
1174
1175         err = -EINVAL;
1176
1177         if (sb->s_header.h_magic != cpu_to_be32(JBD2_MAGIC_NUMBER) ||
1178             sb->s_blocksize != cpu_to_be32(journal->j_blocksize)) {
1179                 printk(KERN_WARNING "JBD: no valid journal superblock found\n");
1180                 goto out;
1181         }
1182
1183         switch(be32_to_cpu(sb->s_header.h_blocktype)) {
1184         case JBD2_SUPERBLOCK_V1:
1185                 journal->j_format_version = 1;
1186                 break;
1187         case JBD2_SUPERBLOCK_V2:
1188                 journal->j_format_version = 2;
1189                 break;
1190         default:
1191                 printk(KERN_WARNING "JBD: unrecognised superblock format ID\n");
1192                 goto out;
1193         }
1194
1195         if (be32_to_cpu(sb->s_maxlen) < journal->j_maxlen)
1196                 journal->j_maxlen = be32_to_cpu(sb->s_maxlen);
1197         else if (be32_to_cpu(sb->s_maxlen) > journal->j_maxlen) {
1198                 printk (KERN_WARNING "JBD: journal file too short\n");
1199                 goto out;
1200         }
1201
1202         return 0;
1203
1204 out:
1205         journal_fail_superblock(journal);
1206         return err;
1207 }
1208
1209 /*
1210  * Load the on-disk journal superblock and read the key fields into the
1211  * journal_t.
1212  */
1213
1214 static int load_superblock(journal_t *journal)
1215 {
1216         int err;
1217         journal_superblock_t *sb;
1218
1219         err = journal_get_superblock(journal);
1220         if (err)
1221                 return err;
1222
1223         sb = journal->j_superblock;
1224
1225         journal->j_tail_sequence = be32_to_cpu(sb->s_sequence);
1226         journal->j_tail = be32_to_cpu(sb->s_start);
1227         journal->j_first = be32_to_cpu(sb->s_first);
1228         journal->j_last = be32_to_cpu(sb->s_maxlen);
1229         journal->j_errno = be32_to_cpu(sb->s_errno);
1230
1231         return 0;
1232 }
1233
1234
1235 /**
1236  * int jbd2_journal_load() - Read journal from disk.
1237  * @journal: Journal to act on.
1238  *
1239  * Given a journal_t structure which tells us which disk blocks contain
1240  * a journal, read the journal from disk to initialise the in-memory
1241  * structures.
1242  */
1243 int jbd2_journal_load(journal_t *journal)
1244 {
1245         int err;
1246         journal_superblock_t *sb;
1247
1248         err = load_superblock(journal);
1249         if (err)
1250                 return err;
1251
1252         sb = journal->j_superblock;
1253         /* If this is a V2 superblock, then we have to check the
1254          * features flags on it. */
1255
1256         if (journal->j_format_version >= 2) {
1257                 if ((sb->s_feature_ro_compat &
1258                      ~cpu_to_be32(JBD2_KNOWN_ROCOMPAT_FEATURES)) ||
1259                     (sb->s_feature_incompat &
1260                      ~cpu_to_be32(JBD2_KNOWN_INCOMPAT_FEATURES))) {
1261                         printk (KERN_WARNING
1262                                 "JBD: Unrecognised features on journal\n");
1263                         return -EINVAL;
1264                 }
1265         }
1266
1267         /*
1268          * Create a slab for this blocksize
1269          */
1270         err = jbd2_journal_create_slab(be32_to_cpu(sb->s_blocksize));
1271         if (err)
1272                 return err;
1273
1274         /* Let the recovery code check whether it needs to recover any
1275          * data from the journal. */
1276         if (jbd2_journal_recover(journal))
1277                 goto recovery_error;
1278
1279         if (journal->j_failed_commit) {
1280                 printk(KERN_ERR "JBD2: journal transaction %u on %s "
1281                        "is corrupt.\n", journal->j_failed_commit,
1282                        journal->j_devname);
1283                 return -EIO;
1284         }
1285
1286         /* OK, we've finished with the dynamic journal bits:
1287          * reinitialise the dynamic contents of the superblock in memory
1288          * and reset them on disk. */
1289         if (journal_reset(journal))
1290                 goto recovery_error;
1291
1292         journal->j_flags &= ~JBD2_ABORT;
1293         journal->j_flags |= JBD2_LOADED;
1294         return 0;
1295
1296 recovery_error:
1297         printk (KERN_WARNING "JBD: recovery failed\n");
1298         return -EIO;
1299 }
1300
1301 /**
1302  * void jbd2_journal_destroy() - Release a journal_t structure.
1303  * @journal: Journal to act on.
1304  *
1305  * Release a journal_t structure once it is no longer in use by the
1306  * journaled object.
1307  * Return <0 if we couldn't clean up the journal.
1308  */
1309 int jbd2_journal_destroy(journal_t *journal)
1310 {
1311         int err = 0;
1312
1313         /* Wait for the commit thread to wake up and die. */
1314         journal_kill_thread(journal);
1315
1316         /* Force a final log commit */
1317         if (journal->j_running_transaction)
1318                 jbd2_journal_commit_transaction(journal);
1319
1320         /* Force any old transactions to disk */
1321
1322         /* Totally anal locking here... */
1323         spin_lock(&journal->j_list_lock);
1324         while (journal->j_checkpoint_transactions != NULL) {
1325                 spin_unlock(&journal->j_list_lock);
1326                 mutex_lock(&journal->j_checkpoint_mutex);
1327                 jbd2_log_do_checkpoint(journal);
1328                 mutex_unlock(&journal->j_checkpoint_mutex);
1329                 spin_lock(&journal->j_list_lock);
1330         }
1331
1332         J_ASSERT(journal->j_running_transaction == NULL);
1333         J_ASSERT(journal->j_committing_transaction == NULL);
1334         J_ASSERT(journal->j_checkpoint_transactions == NULL);
1335         spin_unlock(&journal->j_list_lock);
1336
1337         if (journal->j_sb_buffer) {
1338                 if (!is_journal_aborted(journal)) {
1339                         /* We can now mark the journal as empty. */
1340                         journal->j_tail = 0;
1341                         journal->j_tail_sequence =
1342                                 ++journal->j_transaction_sequence;
1343                         jbd2_journal_update_superblock(journal, 1);
1344                 } else {
1345                         err = -EIO;
1346                 }
1347                 brelse(journal->j_sb_buffer);
1348         }
1349
1350         if (journal->j_proc_entry)
1351                 jbd2_stats_proc_exit(journal);
1352         if (journal->j_inode)
1353                 iput(journal->j_inode);
1354         if (journal->j_revoke)
1355                 jbd2_journal_destroy_revoke(journal);
1356         kfree(journal->j_wbuf);
1357         kfree(journal);
1358
1359         return err;
1360 }
1361
1362
1363 /**
1364  *int jbd2_journal_check_used_features () - Check if features specified are used.
1365  * @journal: Journal to check.
1366  * @compat: bitmask of compatible features
1367  * @ro: bitmask of features that force read-only mount
1368  * @incompat: bitmask of incompatible features
1369  *
1370  * Check whether the journal uses all of a given set of
1371  * features.  Return true (non-zero) if it does.
1372  **/
1373
1374 int jbd2_journal_check_used_features (journal_t *journal, unsigned long compat,
1375                                  unsigned long ro, unsigned long incompat)
1376 {
1377         journal_superblock_t *sb;
1378
1379         if (!compat && !ro && !incompat)
1380                 return 1;
1381         /* Load journal superblock if it is not loaded yet. */
1382         if (journal->j_format_version == 0 &&
1383             journal_get_superblock(journal) != 0)
1384                 return 0;
1385         if (journal->j_format_version == 1)
1386                 return 0;
1387
1388         sb = journal->j_superblock;
1389
1390         if (((be32_to_cpu(sb->s_feature_compat) & compat) == compat) &&
1391             ((be32_to_cpu(sb->s_feature_ro_compat) & ro) == ro) &&
1392             ((be32_to_cpu(sb->s_feature_incompat) & incompat) == incompat))
1393                 return 1;
1394
1395         return 0;
1396 }
1397
1398 /**
1399  * int jbd2_journal_check_available_features() - Check feature set in journalling layer
1400  * @journal: Journal to check.
1401  * @compat: bitmask of compatible features
1402  * @ro: bitmask of features that force read-only mount
1403  * @incompat: bitmask of incompatible features
1404  *
1405  * Check whether the journaling code supports the use of
1406  * all of a given set of features on this journal.  Return true
1407  * (non-zero) if it can. */
1408
1409 int jbd2_journal_check_available_features (journal_t *journal, unsigned long compat,
1410                                       unsigned long ro, unsigned long incompat)
1411 {
1412         if (!compat && !ro && !incompat)
1413                 return 1;
1414
1415         /* We can support any known requested features iff the
1416          * superblock is in version 2.  Otherwise we fail to support any
1417          * extended sb features. */
1418
1419         if (journal->j_format_version != 2)
1420                 return 0;
1421
1422         if ((compat   & JBD2_KNOWN_COMPAT_FEATURES) == compat &&
1423             (ro       & JBD2_KNOWN_ROCOMPAT_FEATURES) == ro &&
1424             (incompat & JBD2_KNOWN_INCOMPAT_FEATURES) == incompat)
1425                 return 1;
1426
1427         return 0;
1428 }
1429
1430 /**
1431  * int jbd2_journal_set_features () - Mark a given journal feature in the superblock
1432  * @journal: Journal to act on.
1433  * @compat: bitmask of compatible features
1434  * @ro: bitmask of features that force read-only mount
1435  * @incompat: bitmask of incompatible features
1436  *
1437  * Mark a given journal feature as present on the
1438  * superblock.  Returns true if the requested features could be set.
1439  *
1440  */
1441
1442 int jbd2_journal_set_features (journal_t *journal, unsigned long compat,
1443                           unsigned long ro, unsigned long incompat)
1444 {
1445         journal_superblock_t *sb;
1446
1447         if (jbd2_journal_check_used_features(journal, compat, ro, incompat))
1448                 return 1;
1449
1450         if (!jbd2_journal_check_available_features(journal, compat, ro, incompat))
1451                 return 0;
1452
1453         jbd_debug(1, "Setting new features 0x%lx/0x%lx/0x%lx\n",
1454                   compat, ro, incompat);
1455
1456         sb = journal->j_superblock;
1457
1458         sb->s_feature_compat    |= cpu_to_be32(compat);
1459         sb->s_feature_ro_compat |= cpu_to_be32(ro);
1460         sb->s_feature_incompat  |= cpu_to_be32(incompat);
1461
1462         return 1;
1463 }
1464
1465 /*
1466  * jbd2_journal_clear_features () - Clear a given journal feature in the
1467  *                                  superblock
1468  * @journal: Journal to act on.
1469  * @compat: bitmask of compatible features
1470  * @ro: bitmask of features that force read-only mount
1471  * @incompat: bitmask of incompatible features
1472  *
1473  * Clear a given journal feature as present on the
1474  * superblock.
1475  */
1476 void jbd2_journal_clear_features(journal_t *journal, unsigned long compat,
1477                                 unsigned long ro, unsigned long incompat)
1478 {
1479         journal_superblock_t *sb;
1480
1481         jbd_debug(1, "Clear features 0x%lx/0x%lx/0x%lx\n",
1482                   compat, ro, incompat);
1483
1484         sb = journal->j_superblock;
1485
1486         sb->s_feature_compat    &= ~cpu_to_be32(compat);
1487         sb->s_feature_ro_compat &= ~cpu_to_be32(ro);
1488         sb->s_feature_incompat  &= ~cpu_to_be32(incompat);
1489 }
1490 EXPORT_SYMBOL(jbd2_journal_clear_features);
1491
1492 /**
1493  * int jbd2_journal_update_format () - Update on-disk journal structure.
1494  * @journal: Journal to act on.
1495  *
1496  * Given an initialised but unloaded journal struct, poke about in the
1497  * on-disk structure to update it to the most recent supported version.
1498  */
1499 int jbd2_journal_update_format (journal_t *journal)
1500 {
1501         journal_superblock_t *sb;
1502         int err;
1503
1504         err = journal_get_superblock(journal);
1505         if (err)
1506                 return err;
1507
1508         sb = journal->j_superblock;
1509
1510         switch (be32_to_cpu(sb->s_header.h_blocktype)) {
1511         case JBD2_SUPERBLOCK_V2:
1512                 return 0;
1513         case JBD2_SUPERBLOCK_V1:
1514                 return journal_convert_superblock_v1(journal, sb);
1515         default:
1516                 break;
1517         }
1518         return -EINVAL;
1519 }
1520
1521 static int journal_convert_superblock_v1(journal_t *journal,
1522                                          journal_superblock_t *sb)
1523 {
1524         int offset, blocksize;
1525         struct buffer_head *bh;
1526
1527         printk(KERN_WARNING
1528                 "JBD: Converting superblock from version 1 to 2.\n");
1529
1530         /* Pre-initialise new fields to zero */
1531         offset = ((char *) &(sb->s_feature_compat)) - ((char *) sb);
1532         blocksize = be32_to_cpu(sb->s_blocksize);
1533         memset(&sb->s_feature_compat, 0, blocksize-offset);
1534
1535         sb->s_nr_users = cpu_to_be32(1);
1536         sb->s_header.h_blocktype = cpu_to_be32(JBD2_SUPERBLOCK_V2);
1537         journal->j_format_version = 2;
1538
1539         bh = journal->j_sb_buffer;
1540         BUFFER_TRACE(bh, "marking dirty");
1541         mark_buffer_dirty(bh);
1542         sync_dirty_buffer(bh);
1543         return 0;
1544 }
1545
1546
1547 /**
1548  * int jbd2_journal_flush () - Flush journal
1549  * @journal: Journal to act on.
1550  *
1551  * Flush all data for a given journal to disk and empty the journal.
1552  * Filesystems can use this when remounting readonly to ensure that
1553  * recovery does not need to happen on remount.
1554  */
1555
1556 int jbd2_journal_flush(journal_t *journal)
1557 {
1558         int err = 0;
1559         transaction_t *transaction = NULL;
1560         unsigned long old_tail;
1561
1562         write_lock(&journal->j_state_lock);
1563
1564         /* Force everything buffered to the log... */
1565         if (journal->j_running_transaction) {
1566                 transaction = journal->j_running_transaction;
1567                 __jbd2_log_start_commit(journal, transaction->t_tid);
1568         } else if (journal->j_committing_transaction)
1569                 transaction = journal->j_committing_transaction;
1570
1571         /* Wait for the log commit to complete... */
1572         if (transaction) {
1573                 tid_t tid = transaction->t_tid;
1574
1575                 write_unlock(&journal->j_state_lock);
1576                 jbd2_log_wait_commit(journal, tid);
1577         } else {
1578                 write_unlock(&journal->j_state_lock);
1579         }
1580
1581         /* ...and flush everything in the log out to disk. */
1582         spin_lock(&journal->j_list_lock);
1583         while (!err && journal->j_checkpoint_transactions != NULL) {
1584                 spin_unlock(&journal->j_list_lock);
1585                 mutex_lock(&journal->j_checkpoint_mutex);
1586                 err = jbd2_log_do_checkpoint(journal);
1587                 mutex_unlock(&journal->j_checkpoint_mutex);
1588                 spin_lock(&journal->j_list_lock);
1589         }
1590         spin_unlock(&journal->j_list_lock);
1591
1592         if (is_journal_aborted(journal))
1593                 return -EIO;
1594
1595         jbd2_cleanup_journal_tail(journal);
1596
1597         /* Finally, mark the journal as really needing no recovery.
1598          * This sets s_start==0 in the underlying superblock, which is
1599          * the magic code for a fully-recovered superblock.  Any future
1600          * commits of data to the journal will restore the current
1601          * s_start value. */
1602         write_lock(&journal->j_state_lock);
1603         old_tail = journal->j_tail;
1604         journal->j_tail = 0;
1605         write_unlock(&journal->j_state_lock);
1606         jbd2_journal_update_superblock(journal, 1);
1607         write_lock(&journal->j_state_lock);
1608         journal->j_tail = old_tail;
1609
1610         J_ASSERT(!journal->j_running_transaction);
1611         J_ASSERT(!journal->j_committing_transaction);
1612         J_ASSERT(!journal->j_checkpoint_transactions);
1613         J_ASSERT(journal->j_head == journal->j_tail);
1614         J_ASSERT(journal->j_tail_sequence == journal->j_transaction_sequence);
1615         write_unlock(&journal->j_state_lock);
1616         return 0;
1617 }
1618
1619 /**
1620  * int jbd2_journal_wipe() - Wipe journal contents
1621  * @journal: Journal to act on.
1622  * @write: flag (see below)
1623  *
1624  * Wipe out all of the contents of a journal, safely.  This will produce
1625  * a warning if the journal contains any valid recovery information.
1626  * Must be called between journal_init_*() and jbd2_journal_load().
1627  *
1628  * If 'write' is non-zero, then we wipe out the journal on disk; otherwise
1629  * we merely suppress recovery.
1630  */
1631
1632 int jbd2_journal_wipe(journal_t *journal, int write)
1633 {
1634         int err = 0;
1635
1636         J_ASSERT (!(journal->j_flags & JBD2_LOADED));
1637
1638         err = load_superblock(journal);
1639         if (err)
1640                 return err;
1641
1642         if (!journal->j_tail)
1643                 goto no_recovery;
1644
1645         printk (KERN_WARNING "JBD: %s recovery information on journal\n",
1646                 write ? "Clearing" : "Ignoring");
1647
1648         err = jbd2_journal_skip_recovery(journal);
1649         if (write)
1650                 jbd2_journal_update_superblock(journal, 1);
1651
1652  no_recovery:
1653         return err;
1654 }
1655
1656 /*
1657  * Journal abort has very specific semantics, which we describe
1658  * for journal abort.
1659  *
1660  * Two internal functions, which provide abort to the jbd layer
1661  * itself are here.
1662  */
1663
1664 /*
1665  * Quick version for internal journal use (doesn't lock the journal).
1666  * Aborts hard --- we mark the abort as occurred, but do _nothing_ else,
1667  * and don't attempt to make any other journal updates.
1668  */
1669 void __jbd2_journal_abort_hard(journal_t *journal)
1670 {
1671         transaction_t *transaction;
1672
1673         if (journal->j_flags & JBD2_ABORT)
1674                 return;
1675
1676         printk(KERN_ERR "Aborting journal on device %s.\n",
1677                journal->j_devname);
1678
1679         write_lock(&journal->j_state_lock);
1680         journal->j_flags |= JBD2_ABORT;
1681         transaction = journal->j_running_transaction;
1682         if (transaction)
1683                 __jbd2_log_start_commit(journal, transaction->t_tid);
1684         write_unlock(&journal->j_state_lock);
1685 }
1686
1687 /* Soft abort: record the abort error status in the journal superblock,
1688  * but don't do any other IO. */
1689 static void __journal_abort_soft (journal_t *journal, int errno)
1690 {
1691         if (journal->j_flags & JBD2_ABORT)
1692                 return;
1693
1694         if (!journal->j_errno)
1695                 journal->j_errno = errno;
1696
1697         __jbd2_journal_abort_hard(journal);
1698
1699         if (errno)
1700                 jbd2_journal_update_superblock(journal, 1);
1701 }
1702
1703 /**
1704  * void jbd2_journal_abort () - Shutdown the journal immediately.
1705  * @journal: the journal to shutdown.
1706  * @errno:   an error number to record in the journal indicating
1707  *           the reason for the shutdown.
1708  *
1709  * Perform a complete, immediate shutdown of the ENTIRE
1710  * journal (not of a single transaction).  This operation cannot be
1711  * undone without closing and reopening the journal.
1712  *
1713  * The jbd2_journal_abort function is intended to support higher level error
1714  * recovery mechanisms such as the ext2/ext3 remount-readonly error
1715  * mode.
1716  *
1717  * Journal abort has very specific semantics.  Any existing dirty,
1718  * unjournaled buffers in the main filesystem will still be written to
1719  * disk by bdflush, but the journaling mechanism will be suspended
1720  * immediately and no further transaction commits will be honoured.
1721  *
1722  * Any dirty, journaled buffers will be written back to disk without
1723  * hitting the journal.  Atomicity cannot be guaranteed on an aborted
1724  * filesystem, but we _do_ attempt to leave as much data as possible
1725  * behind for fsck to use for cleanup.
1726  *
1727  * Any attempt to get a new transaction handle on a journal which is in
1728  * ABORT state will just result in an -EROFS error return.  A
1729  * jbd2_journal_stop on an existing handle will return -EIO if we have
1730  * entered abort state during the update.
1731  *
1732  * Recursive transactions are not disturbed by journal abort until the
1733  * final jbd2_journal_stop, which will receive the -EIO error.
1734  *
1735  * Finally, the jbd2_journal_abort call allows the caller to supply an errno
1736  * which will be recorded (if possible) in the journal superblock.  This
1737  * allows a client to record failure conditions in the middle of a
1738  * transaction without having to complete the transaction to record the
1739  * failure to disk.  ext3_error, for example, now uses this
1740  * functionality.
1741  *
1742  * Errors which originate from within the journaling layer will NOT
1743  * supply an errno; a null errno implies that absolutely no further
1744  * writes are done to the journal (unless there are any already in
1745  * progress).
1746  *
1747  */
1748
1749 void jbd2_journal_abort(journal_t *journal, int errno)
1750 {
1751         __journal_abort_soft(journal, errno);
1752 }
1753
1754 /**
1755  * int jbd2_journal_errno () - returns the journal's error state.
1756  * @journal: journal to examine.
1757  *
1758  * This is the errno number set with jbd2_journal_abort(), the last
1759  * time the journal was mounted - if the journal was stopped
1760  * without calling abort this will be 0.
1761  *
1762  * If the journal has been aborted on this mount time -EROFS will
1763  * be returned.
1764  */
1765 int jbd2_journal_errno(journal_t *journal)
1766 {
1767         int err;
1768
1769         read_lock(&journal->j_state_lock);
1770         if (journal->j_flags & JBD2_ABORT)
1771                 err = -EROFS;
1772         else
1773                 err = journal->j_errno;
1774         read_unlock(&journal->j_state_lock);
1775         return err;
1776 }
1777
1778 /**
1779  * int jbd2_journal_clear_err () - clears the journal's error state
1780  * @journal: journal to act on.
1781  *
1782  * An error must be cleared or acked to take a FS out of readonly
1783  * mode.
1784  */
1785 int jbd2_journal_clear_err(journal_t *journal)
1786 {
1787         int err = 0;
1788
1789         write_lock(&journal->j_state_lock);
1790         if (journal->j_flags & JBD2_ABORT)
1791                 err = -EROFS;
1792         else
1793                 journal->j_errno = 0;
1794         write_unlock(&journal->j_state_lock);
1795         return err;
1796 }
1797
1798 /**
1799  * void jbd2_journal_ack_err() - Ack journal err.
1800  * @journal: journal to act on.
1801  *
1802  * An error must be cleared or acked to take a FS out of readonly
1803  * mode.
1804  */
1805 void jbd2_journal_ack_err(journal_t *journal)
1806 {
1807         write_lock(&journal->j_state_lock);
1808         if (journal->j_errno)
1809                 journal->j_flags |= JBD2_ACK_ERR;
1810         write_unlock(&journal->j_state_lock);
1811 }
1812
1813 int jbd2_journal_blocks_per_page(struct inode *inode)
1814 {
1815         return 1 << (PAGE_CACHE_SHIFT - inode->i_sb->s_blocksize_bits);
1816 }
1817
1818 /*
1819  * helper functions to deal with 32 or 64bit block numbers.
1820  */
1821 size_t journal_tag_bytes(journal_t *journal)
1822 {
1823         if (JBD2_HAS_INCOMPAT_FEATURE(journal, JBD2_FEATURE_INCOMPAT_64BIT))
1824                 return JBD2_TAG_SIZE64;
1825         else
1826                 return JBD2_TAG_SIZE32;
1827 }
1828
1829 /*
1830  * JBD memory management
1831  *
1832  * These functions are used to allocate block-sized chunks of memory
1833  * used for making copies of buffer_head data.  Very often it will be
1834  * page-sized chunks of data, but sometimes it will be in
1835  * sub-page-size chunks.  (For example, 16k pages on Power systems
1836  * with a 4k block file system.)  For blocks smaller than a page, we
1837  * use a SLAB allocator.  There are slab caches for each block size,
1838  * which are allocated at mount time, if necessary, and we only free
1839  * (all of) the slab caches when/if the jbd2 module is unloaded.  For
1840  * this reason we don't need to a mutex to protect access to
1841  * jbd2_slab[] allocating or releasing memory; only in
1842  * jbd2_journal_create_slab().
1843  */
1844 #define JBD2_MAX_SLABS 8
1845 static struct kmem_cache *jbd2_slab[JBD2_MAX_SLABS];
1846
1847 static const char *jbd2_slab_names[JBD2_MAX_SLABS] = {
1848         "jbd2_1k", "jbd2_2k", "jbd2_4k", "jbd2_8k",
1849         "jbd2_16k", "jbd2_32k", "jbd2_64k", "jbd2_128k"
1850 };
1851
1852
1853 static void jbd2_journal_destroy_slabs(void)
1854 {
1855         int i;
1856
1857         for (i = 0; i < JBD2_MAX_SLABS; i++) {
1858                 if (jbd2_slab[i])
1859                         kmem_cache_destroy(jbd2_slab[i]);
1860                 jbd2_slab[i] = NULL;
1861         }
1862 }
1863
1864 static int jbd2_journal_create_slab(size_t size)
1865 {
1866         static DEFINE_MUTEX(jbd2_slab_create_mutex);
1867         int i = order_base_2(size) - 10;
1868         size_t slab_size;
1869
1870         if (size == PAGE_SIZE)
1871                 return 0;
1872
1873         if (i >= JBD2_MAX_SLABS)
1874                 return -EINVAL;
1875
1876         if (unlikely(i < 0))
1877                 i = 0;
1878         mutex_lock(&jbd2_slab_create_mutex);
1879         if (jbd2_slab[i]) {
1880                 mutex_unlock(&jbd2_slab_create_mutex);
1881                 return 0;       /* Already created */
1882         }
1883
1884         slab_size = 1 << (i+10);
1885         jbd2_slab[i] = kmem_cache_create(jbd2_slab_names[i], slab_size,
1886                                          slab_size, 0, NULL);
1887         mutex_unlock(&jbd2_slab_create_mutex);
1888         if (!jbd2_slab[i]) {
1889                 printk(KERN_EMERG "JBD2: no memory for jbd2_slab cache\n");
1890                 return -ENOMEM;
1891         }
1892         return 0;
1893 }
1894
1895 static struct kmem_cache *get_slab(size_t size)
1896 {
1897         int i = order_base_2(size) - 10;
1898
1899         BUG_ON(i >= JBD2_MAX_SLABS);
1900         if (unlikely(i < 0))
1901                 i = 0;
1902         BUG_ON(jbd2_slab[i] == NULL);
1903         return jbd2_slab[i];
1904 }
1905
1906 void *jbd2_alloc(size_t size, gfp_t flags)
1907 {
1908         void *ptr;
1909
1910         BUG_ON(size & (size-1)); /* Must be a power of 2 */
1911
1912         flags |= __GFP_REPEAT;
1913         if (size == PAGE_SIZE)
1914                 ptr = (void *)__get_free_pages(flags, 0);
1915         else if (size > PAGE_SIZE) {
1916                 int order = get_order(size);
1917
1918                 if (order < 3)
1919                         ptr = (void *)__get_free_pages(flags, order);
1920                 else
1921                         ptr = vmalloc(size);
1922         } else
1923                 ptr = kmem_cache_alloc(get_slab(size), flags);
1924
1925         /* Check alignment; SLUB has gotten this wrong in the past,
1926          * and this can lead to user data corruption! */
1927         BUG_ON(((unsigned long) ptr) & (size-1));
1928
1929         return ptr;
1930 }
1931
1932 void jbd2_free(void *ptr, size_t size)
1933 {
1934         if (size == PAGE_SIZE) {
1935                 free_pages((unsigned long)ptr, 0);
1936                 return;
1937         }
1938         if (size > PAGE_SIZE) {
1939                 int order = get_order(size);
1940
1941                 if (order < 3)
1942                         free_pages((unsigned long)ptr, order);
1943                 else
1944                         vfree(ptr);
1945                 return;
1946         }
1947         kmem_cache_free(get_slab(size), ptr);
1948 };
1949
1950 /*
1951  * Journal_head storage management
1952  */
1953 static struct kmem_cache *jbd2_journal_head_cache;
1954 #ifdef CONFIG_JBD2_DEBUG
1955 static atomic_t nr_journal_heads = ATOMIC_INIT(0);
1956 #endif
1957
1958 static int journal_init_jbd2_journal_head_cache(void)
1959 {
1960         int retval;
1961
1962         J_ASSERT(jbd2_journal_head_cache == NULL);
1963         jbd2_journal_head_cache = kmem_cache_create("jbd2_journal_head",
1964                                 sizeof(struct journal_head),
1965                                 0,              /* offset */
1966                                 SLAB_TEMPORARY, /* flags */
1967                                 NULL);          /* ctor */
1968         retval = 0;
1969         if (!jbd2_journal_head_cache) {
1970                 retval = -ENOMEM;
1971                 printk(KERN_EMERG "JBD: no memory for journal_head cache\n");
1972         }
1973         return retval;
1974 }
1975
1976 static void jbd2_journal_destroy_jbd2_journal_head_cache(void)
1977 {
1978         if (jbd2_journal_head_cache) {
1979                 kmem_cache_destroy(jbd2_journal_head_cache);
1980                 jbd2_journal_head_cache = NULL;
1981         }
1982 }
1983
1984 /*
1985  * journal_head splicing and dicing
1986  */
1987 static struct journal_head *journal_alloc_journal_head(void)
1988 {
1989         struct journal_head *ret;
1990
1991 #ifdef CONFIG_JBD2_DEBUG
1992         atomic_inc(&nr_journal_heads);
1993 #endif
1994         ret = kmem_cache_alloc(jbd2_journal_head_cache, GFP_NOFS);
1995         if (!ret) {
1996                 jbd_debug(1, "out of memory for journal_head\n");
1997                 pr_notice_ratelimited("ENOMEM in %s, retrying.\n", __func__);
1998                 while (!ret) {
1999                         yield();
2000                         ret = kmem_cache_alloc(jbd2_journal_head_cache, GFP_NOFS);
2001                 }
2002         }
2003         return ret;
2004 }
2005
2006 static void journal_free_journal_head(struct journal_head *jh)
2007 {
2008 #ifdef CONFIG_JBD2_DEBUG
2009         atomic_dec(&nr_journal_heads);
2010         memset(jh, JBD2_POISON_FREE, sizeof(*jh));
2011 #endif
2012         kmem_cache_free(jbd2_journal_head_cache, jh);
2013 }
2014
2015 /*
2016  * A journal_head is attached to a buffer_head whenever JBD has an
2017  * interest in the buffer.
2018  *
2019  * Whenever a buffer has an attached journal_head, its ->b_state:BH_JBD bit
2020  * is set.  This bit is tested in core kernel code where we need to take
2021  * JBD-specific actions.  Testing the zeroness of ->b_private is not reliable
2022  * there.
2023  *
2024  * When a buffer has its BH_JBD bit set, its ->b_count is elevated by one.
2025  *
2026  * When a buffer has its BH_JBD bit set it is immune from being released by
2027  * core kernel code, mainly via ->b_count.
2028  *
2029  * A journal_head may be detached from its buffer_head when the journal_head's
2030  * b_transaction, b_cp_transaction and b_next_transaction pointers are NULL.
2031  * Various places in JBD call jbd2_journal_remove_journal_head() to indicate that the
2032  * journal_head can be dropped if needed.
2033  *
2034  * Various places in the kernel want to attach a journal_head to a buffer_head
2035  * _before_ attaching the journal_head to a transaction.  To protect the
2036  * journal_head in this situation, jbd2_journal_add_journal_head elevates the
2037  * journal_head's b_jcount refcount by one.  The caller must call
2038  * jbd2_journal_put_journal_head() to undo this.
2039  *
2040  * So the typical usage would be:
2041  *
2042  *      (Attach a journal_head if needed.  Increments b_jcount)
2043  *      struct journal_head *jh = jbd2_journal_add_journal_head(bh);
2044  *      ...
2045  *      jh->b_transaction = xxx;
2046  *      jbd2_journal_put_journal_head(jh);
2047  *
2048  * Now, the journal_head's b_jcount is zero, but it is safe from being released
2049  * because it has a non-zero b_transaction.
2050  */
2051
2052 /*
2053  * Give a buffer_head a journal_head.
2054  *
2055  * Doesn't need the journal lock.
2056  * May sleep.
2057  */
2058 struct journal_head *jbd2_journal_add_journal_head(struct buffer_head *bh)
2059 {
2060         struct journal_head *jh;
2061         struct journal_head *new_jh = NULL;
2062
2063 repeat:
2064         if (!buffer_jbd(bh)) {
2065                 new_jh = journal_alloc_journal_head();
2066                 memset(new_jh, 0, sizeof(*new_jh));
2067         }
2068
2069         jbd_lock_bh_journal_head(bh);
2070         if (buffer_jbd(bh)) {
2071                 jh = bh2jh(bh);
2072         } else {
2073                 J_ASSERT_BH(bh,
2074                         (atomic_read(&bh->b_count) > 0) ||
2075                         (bh->b_page && bh->b_page->mapping));
2076
2077                 if (!new_jh) {
2078                         jbd_unlock_bh_journal_head(bh);
2079                         goto repeat;
2080                 }
2081
2082                 jh = new_jh;
2083                 new_jh = NULL;          /* We consumed it */
2084                 set_buffer_jbd(bh);
2085                 bh->b_private = jh;
2086                 jh->b_bh = bh;
2087                 get_bh(bh);
2088                 BUFFER_TRACE(bh, "added journal_head");
2089         }
2090         jh->b_jcount++;
2091         jbd_unlock_bh_journal_head(bh);
2092         if (new_jh)
2093                 journal_free_journal_head(new_jh);
2094         return bh->b_private;
2095 }
2096
2097 /*
2098  * Grab a ref against this buffer_head's journal_head.  If it ended up not
2099  * having a journal_head, return NULL
2100  */
2101 struct journal_head *jbd2_journal_grab_journal_head(struct buffer_head *bh)
2102 {
2103         struct journal_head *jh = NULL;
2104
2105         jbd_lock_bh_journal_head(bh);
2106         if (buffer_jbd(bh)) {
2107                 jh = bh2jh(bh);
2108                 jh->b_jcount++;
2109         }
2110         jbd_unlock_bh_journal_head(bh);
2111         return jh;
2112 }
2113
2114 static void __journal_remove_journal_head(struct buffer_head *bh)
2115 {
2116         struct journal_head *jh = bh2jh(bh);
2117
2118         J_ASSERT_JH(jh, jh->b_jcount >= 0);
2119
2120         get_bh(bh);
2121         if (jh->b_jcount == 0) {
2122                 if (jh->b_transaction == NULL &&
2123                                 jh->b_next_transaction == NULL &&
2124                                 jh->b_cp_transaction == NULL) {
2125                         J_ASSERT_JH(jh, jh->b_jlist == BJ_None);
2126                         J_ASSERT_BH(bh, buffer_jbd(bh));
2127                         J_ASSERT_BH(bh, jh2bh(jh) == bh);
2128                         BUFFER_TRACE(bh, "remove journal_head");
2129                         if (jh->b_frozen_data) {
2130                                 printk(KERN_WARNING "%s: freeing "
2131                                                 "b_frozen_data\n",
2132                                                 __func__);
2133                                 jbd2_free(jh->b_frozen_data, bh->b_size);
2134                         }
2135                         if (jh->b_committed_data) {
2136                                 printk(KERN_WARNING "%s: freeing "
2137                                                 "b_committed_data\n",
2138                                                 __func__);
2139                                 jbd2_free(jh->b_committed_data, bh->b_size);
2140                         }
2141                         bh->b_private = NULL;
2142                         jh->b_bh = NULL;        /* debug, really */
2143                         clear_buffer_jbd(bh);
2144                         __brelse(bh);
2145                         journal_free_journal_head(jh);
2146                 } else {
2147                         BUFFER_TRACE(bh, "journal_head was locked");
2148                 }
2149         }
2150 }
2151
2152 /*
2153  * jbd2_journal_remove_journal_head(): if the buffer isn't attached to a transaction
2154  * and has a zero b_jcount then remove and release its journal_head.   If we did
2155  * see that the buffer is not used by any transaction we also "logically"
2156  * decrement ->b_count.
2157  *
2158  * We in fact take an additional increment on ->b_count as a convenience,
2159  * because the caller usually wants to do additional things with the bh
2160  * after calling here.
2161  * The caller of jbd2_journal_remove_journal_head() *must* run __brelse(bh) at some
2162  * time.  Once the caller has run __brelse(), the buffer is eligible for
2163  * reaping by try_to_free_buffers().
2164  */
2165 void jbd2_journal_remove_journal_head(struct buffer_head *bh)
2166 {
2167         jbd_lock_bh_journal_head(bh);
2168         __journal_remove_journal_head(bh);
2169         jbd_unlock_bh_journal_head(bh);
2170 }
2171
2172 /*
2173  * Drop a reference on the passed journal_head.  If it fell to zero then try to
2174  * release the journal_head from the buffer_head.
2175  */
2176 void jbd2_journal_put_journal_head(struct journal_head *jh)
2177 {
2178         struct buffer_head *bh = jh2bh(jh);
2179
2180         jbd_lock_bh_journal_head(bh);
2181         J_ASSERT_JH(jh, jh->b_jcount > 0);
2182         --jh->b_jcount;
2183         if (!jh->b_jcount && !jh->b_transaction) {
2184                 __journal_remove_journal_head(bh);
2185                 __brelse(bh);
2186         }
2187         jbd_unlock_bh_journal_head(bh);
2188 }
2189
2190 /*
2191  * Initialize jbd inode head
2192  */
2193 void jbd2_journal_init_jbd_inode(struct jbd2_inode *jinode, struct inode *inode)
2194 {
2195         jinode->i_transaction = NULL;
2196         jinode->i_next_transaction = NULL;
2197         jinode->i_vfs_inode = inode;
2198         jinode->i_flags = 0;
2199         INIT_LIST_HEAD(&jinode->i_list);
2200 }
2201
2202 /*
2203  * Function to be called before we start removing inode from memory (i.e.,
2204  * clear_inode() is a fine place to be called from). It removes inode from
2205  * transaction's lists.
2206  */
2207 void jbd2_journal_release_jbd_inode(journal_t *journal,
2208                                     struct jbd2_inode *jinode)
2209 {
2210         if (!journal)
2211                 return;
2212 restart:
2213         spin_lock(&journal->j_list_lock);
2214         /* Is commit writing out inode - we have to wait */
2215         if (test_bit(__JI_COMMIT_RUNNING, &jinode->i_flags)) {
2216                 wait_queue_head_t *wq;
2217                 DEFINE_WAIT_BIT(wait, &jinode->i_flags, __JI_COMMIT_RUNNING);
2218                 wq = bit_waitqueue(&jinode->i_flags, __JI_COMMIT_RUNNING);
2219                 prepare_to_wait(wq, &wait.wait, TASK_UNINTERRUPTIBLE);
2220                 spin_unlock(&journal->j_list_lock);
2221                 schedule();
2222                 finish_wait(wq, &wait.wait);
2223                 goto restart;
2224         }
2225
2226         if (jinode->i_transaction) {
2227                 list_del(&jinode->i_list);
2228                 jinode->i_transaction = NULL;
2229         }
2230         spin_unlock(&journal->j_list_lock);
2231 }
2232
2233 /*
2234  * debugfs tunables
2235  */
2236 #ifdef CONFIG_JBD2_DEBUG
2237 u8 jbd2_journal_enable_debug __read_mostly;
2238 EXPORT_SYMBOL(jbd2_journal_enable_debug);
2239
2240 #define JBD2_DEBUG_NAME "jbd2-debug"
2241
2242 static struct dentry *jbd2_debugfs_dir;
2243 static struct dentry *jbd2_debug;
2244
2245 static void __init jbd2_create_debugfs_entry(void)
2246 {
2247         jbd2_debugfs_dir = debugfs_create_dir("jbd2", NULL);
2248         if (jbd2_debugfs_dir)
2249                 jbd2_debug = debugfs_create_u8(JBD2_DEBUG_NAME,
2250                                                S_IRUGO | S_IWUSR,
2251                                                jbd2_debugfs_dir,
2252                                                &jbd2_journal_enable_debug);
2253 }
2254
2255 static void __exit jbd2_remove_debugfs_entry(void)
2256 {
2257         debugfs_remove(jbd2_debug);
2258         debugfs_remove(jbd2_debugfs_dir);
2259 }
2260
2261 #else
2262
2263 static void __init jbd2_create_debugfs_entry(void)
2264 {
2265 }
2266
2267 static void __exit jbd2_remove_debugfs_entry(void)
2268 {
2269 }
2270
2271 #endif
2272
2273 #ifdef CONFIG_PROC_FS
2274
2275 #define JBD2_STATS_PROC_NAME "fs/jbd2"
2276
2277 static void __init jbd2_create_jbd_stats_proc_entry(void)
2278 {
2279         proc_jbd2_stats = proc_mkdir(JBD2_STATS_PROC_NAME, NULL);
2280 }
2281
2282 static void __exit jbd2_remove_jbd_stats_proc_entry(void)
2283 {
2284         if (proc_jbd2_stats)
2285                 remove_proc_entry(JBD2_STATS_PROC_NAME, NULL);
2286 }
2287
2288 #else
2289
2290 #define jbd2_create_jbd_stats_proc_entry() do {} while (0)
2291 #define jbd2_remove_jbd_stats_proc_entry() do {} while (0)
2292
2293 #endif
2294
2295 struct kmem_cache *jbd2_handle_cache, *jbd2_inode_cache;
2296
2297 static int __init journal_init_handle_cache(void)
2298 {
2299         jbd2_handle_cache = KMEM_CACHE(jbd2_journal_handle, SLAB_TEMPORARY);
2300         if (jbd2_handle_cache == NULL) {
2301                 printk(KERN_EMERG "JBD2: failed to create handle cache\n");
2302                 return -ENOMEM;
2303         }
2304         jbd2_inode_cache = KMEM_CACHE(jbd2_inode, 0);
2305         if (jbd2_inode_cache == NULL) {
2306                 printk(KERN_EMERG "JBD2: failed to create inode cache\n");
2307                 kmem_cache_destroy(jbd2_handle_cache);
2308                 return -ENOMEM;
2309         }
2310         return 0;
2311 }
2312
2313 static void jbd2_journal_destroy_handle_cache(void)
2314 {
2315         if (jbd2_handle_cache)
2316                 kmem_cache_destroy(jbd2_handle_cache);
2317         if (jbd2_inode_cache)
2318                 kmem_cache_destroy(jbd2_inode_cache);
2319
2320 }
2321
2322 /*
2323  * Module startup and shutdown
2324  */
2325
2326 static int __init journal_init_caches(void)
2327 {
2328         int ret;
2329
2330         ret = jbd2_journal_init_revoke_caches();
2331         if (ret == 0)
2332                 ret = journal_init_jbd2_journal_head_cache();
2333         if (ret == 0)
2334                 ret = journal_init_handle_cache();
2335         return ret;
2336 }
2337
2338 static void jbd2_journal_destroy_caches(void)
2339 {
2340         jbd2_journal_destroy_revoke_caches();
2341         jbd2_journal_destroy_jbd2_journal_head_cache();
2342         jbd2_journal_destroy_handle_cache();
2343         jbd2_journal_destroy_slabs();
2344 }
2345
2346 static int __init journal_init(void)
2347 {
2348         int ret;
2349
2350         BUILD_BUG_ON(sizeof(struct journal_superblock_s) != 1024);
2351
2352         ret = journal_init_caches();
2353         if (ret == 0) {
2354                 jbd2_create_debugfs_entry();
2355                 jbd2_create_jbd_stats_proc_entry();
2356         } else {
2357                 jbd2_journal_destroy_caches();
2358         }
2359         return ret;
2360 }
2361
2362 static void __exit journal_exit(void)
2363 {
2364 #ifdef CONFIG_JBD2_DEBUG
2365         int n = atomic_read(&nr_journal_heads);
2366         if (n)
2367                 printk(KERN_EMERG "JBD: leaked %d journal_heads!\n", n);
2368 #endif
2369         jbd2_remove_debugfs_entry();
2370         jbd2_remove_jbd_stats_proc_entry();
2371         jbd2_journal_destroy_caches();
2372 }
2373
2374 /* 
2375  * jbd2_dev_to_name is a utility function used by the jbd2 and ext4 
2376  * tracing infrastructure to map a dev_t to a device name.
2377  *
2378  * The caller should use rcu_read_lock() in order to make sure the
2379  * device name stays valid until its done with it.  We use
2380  * rcu_read_lock() as well to make sure we're safe in case the caller
2381  * gets sloppy, and because rcu_read_lock() is cheap and can be safely
2382  * nested.
2383  */
2384 struct devname_cache {
2385         struct rcu_head rcu;
2386         dev_t           device;
2387         char            devname[BDEVNAME_SIZE];
2388 };
2389 #define CACHE_SIZE_BITS 6
2390 static struct devname_cache *devcache[1 << CACHE_SIZE_BITS];
2391 static DEFINE_SPINLOCK(devname_cache_lock);
2392
2393 static void free_devcache(struct rcu_head *rcu)
2394 {
2395         kfree(rcu);
2396 }
2397
2398 const char *jbd2_dev_to_name(dev_t device)
2399 {
2400         int     i = hash_32(device, CACHE_SIZE_BITS);
2401         char    *ret;
2402         struct block_device *bd;
2403         static struct devname_cache *new_dev;
2404
2405         rcu_read_lock();
2406         if (devcache[i] && devcache[i]->device == device) {
2407                 ret = devcache[i]->devname;
2408                 rcu_read_unlock();
2409                 return ret;
2410         }
2411         rcu_read_unlock();
2412
2413         new_dev = kmalloc(sizeof(struct devname_cache), GFP_KERNEL);
2414         if (!new_dev)
2415                 return "NODEV-ALLOCFAILURE"; /* Something non-NULL */
2416         bd = bdget(device);
2417         spin_lock(&devname_cache_lock);
2418         if (devcache[i]) {
2419                 if (devcache[i]->device == device) {
2420                         kfree(new_dev);
2421                         bdput(bd);
2422                         ret = devcache[i]->devname;
2423                         spin_unlock(&devname_cache_lock);
2424                         return ret;
2425                 }
2426                 call_rcu(&devcache[i]->rcu, free_devcache);
2427         }
2428         devcache[i] = new_dev;
2429         devcache[i]->device = device;
2430         if (bd) {
2431                 bdevname(bd, devcache[i]->devname);
2432                 bdput(bd);
2433         } else
2434                 __bdevname(device, devcache[i]->devname);
2435         ret = devcache[i]->devname;
2436         spin_unlock(&devname_cache_lock);
2437         return ret;
2438 }
2439 EXPORT_SYMBOL(jbd2_dev_to_name);
2440
2441 MODULE_LICENSE("GPL");
2442 module_init(journal_init);
2443 module_exit(journal_exit);
2444