Merge tag 'split-asm_system_h-for-linus-20120328' of git://git.kernel.org/pub/scm...
[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
54 EXPORT_SYMBOL(jbd2_journal_extend);
55 EXPORT_SYMBOL(jbd2_journal_stop);
56 EXPORT_SYMBOL(jbd2_journal_lock_updates);
57 EXPORT_SYMBOL(jbd2_journal_unlock_updates);
58 EXPORT_SYMBOL(jbd2_journal_get_write_access);
59 EXPORT_SYMBOL(jbd2_journal_get_create_access);
60 EXPORT_SYMBOL(jbd2_journal_get_undo_access);
61 EXPORT_SYMBOL(jbd2_journal_set_triggers);
62 EXPORT_SYMBOL(jbd2_journal_dirty_metadata);
63 EXPORT_SYMBOL(jbd2_journal_release_buffer);
64 EXPORT_SYMBOL(jbd2_journal_forget);
65 #if 0
66 EXPORT_SYMBOL(journal_sync_buffer);
67 #endif
68 EXPORT_SYMBOL(jbd2_journal_flush);
69 EXPORT_SYMBOL(jbd2_journal_revoke);
70
71 EXPORT_SYMBOL(jbd2_journal_init_dev);
72 EXPORT_SYMBOL(jbd2_journal_init_inode);
73 EXPORT_SYMBOL(jbd2_journal_check_used_features);
74 EXPORT_SYMBOL(jbd2_journal_check_available_features);
75 EXPORT_SYMBOL(jbd2_journal_set_features);
76 EXPORT_SYMBOL(jbd2_journal_load);
77 EXPORT_SYMBOL(jbd2_journal_destroy);
78 EXPORT_SYMBOL(jbd2_journal_abort);
79 EXPORT_SYMBOL(jbd2_journal_errno);
80 EXPORT_SYMBOL(jbd2_journal_ack_err);
81 EXPORT_SYMBOL(jbd2_journal_clear_err);
82 EXPORT_SYMBOL(jbd2_log_wait_commit);
83 EXPORT_SYMBOL(jbd2_log_start_commit);
84 EXPORT_SYMBOL(jbd2_journal_start_commit);
85 EXPORT_SYMBOL(jbd2_journal_force_commit_nested);
86 EXPORT_SYMBOL(jbd2_journal_wipe);
87 EXPORT_SYMBOL(jbd2_journal_blocks_per_page);
88 EXPORT_SYMBOL(jbd2_journal_invalidatepage);
89 EXPORT_SYMBOL(jbd2_journal_try_to_free_buffers);
90 EXPORT_SYMBOL(jbd2_journal_force_commit);
91 EXPORT_SYMBOL(jbd2_journal_file_inode);
92 EXPORT_SYMBOL(jbd2_journal_init_jbd_inode);
93 EXPORT_SYMBOL(jbd2_journal_release_jbd_inode);
94 EXPORT_SYMBOL(jbd2_journal_begin_ordered_truncate);
95 EXPORT_SYMBOL(jbd2_inode_cache);
96
97 static void __journal_abort_soft (journal_t *journal, int errno);
98 static int jbd2_journal_create_slab(size_t slab_size);
99
100 /*
101  * Helper function used to manage commit timeouts
102  */
103
104 static void commit_timeout(unsigned long __data)
105 {
106         struct task_struct * p = (struct task_struct *) __data;
107
108         wake_up_process(p);
109 }
110
111 /*
112  * kjournald2: The main thread function used to manage a logging device
113  * journal.
114  *
115  * This kernel thread is responsible for two things:
116  *
117  * 1) COMMIT:  Every so often we need to commit the current state of the
118  *    filesystem to disk.  The journal thread is responsible for writing
119  *    all of the metadata buffers to disk.
120  *
121  * 2) CHECKPOINT: We cannot reuse a used section of the log file until all
122  *    of the data in that part of the log has been rewritten elsewhere on
123  *    the disk.  Flushing these old buffers to reclaim space in the log is
124  *    known as checkpointing, and this thread is responsible for that job.
125  */
126
127 static int kjournald2(void *arg)
128 {
129         journal_t *journal = arg;
130         transaction_t *transaction;
131
132         /*
133          * Set up an interval timer which can be used to trigger a commit wakeup
134          * after the commit interval expires
135          */
136         setup_timer(&journal->j_commit_timer, commit_timeout,
137                         (unsigned long)current);
138
139         set_freezable();
140
141         /* Record that the journal thread is running */
142         journal->j_task = current;
143         wake_up(&journal->j_wait_done_commit);
144
145         /*
146          * And now, wait forever for commit wakeup events.
147          */
148         write_lock(&journal->j_state_lock);
149
150 loop:
151         if (journal->j_flags & JBD2_UNMOUNT)
152                 goto end_loop;
153
154         jbd_debug(1, "commit_sequence=%d, commit_request=%d\n",
155                 journal->j_commit_sequence, journal->j_commit_request);
156
157         if (journal->j_commit_sequence != journal->j_commit_request) {
158                 jbd_debug(1, "OK, requests differ\n");
159                 write_unlock(&journal->j_state_lock);
160                 del_timer_sync(&journal->j_commit_timer);
161                 jbd2_journal_commit_transaction(journal);
162                 write_lock(&journal->j_state_lock);
163                 goto loop;
164         }
165
166         wake_up(&journal->j_wait_done_commit);
167         if (freezing(current)) {
168                 /*
169                  * The simpler the better. Flushing journal isn't a
170                  * good idea, because that depends on threads that may
171                  * be already stopped.
172                  */
173                 jbd_debug(1, "Now suspending kjournald2\n");
174                 write_unlock(&journal->j_state_lock);
175                 try_to_freeze();
176                 write_lock(&journal->j_state_lock);
177         } else {
178                 /*
179                  * We assume on resume that commits are already there,
180                  * so we don't sleep
181                  */
182                 DEFINE_WAIT(wait);
183                 int should_sleep = 1;
184
185                 prepare_to_wait(&journal->j_wait_commit, &wait,
186                                 TASK_INTERRUPTIBLE);
187                 if (journal->j_commit_sequence != journal->j_commit_request)
188                         should_sleep = 0;
189                 transaction = journal->j_running_transaction;
190                 if (transaction && time_after_eq(jiffies,
191                                                 transaction->t_expires))
192                         should_sleep = 0;
193                 if (journal->j_flags & JBD2_UNMOUNT)
194                         should_sleep = 0;
195                 if (should_sleep) {
196                         write_unlock(&journal->j_state_lock);
197                         schedule();
198                         write_lock(&journal->j_state_lock);
199                 }
200                 finish_wait(&journal->j_wait_commit, &wait);
201         }
202
203         jbd_debug(1, "kjournald2 wakes\n");
204
205         /*
206          * Were we woken up by a commit wakeup event?
207          */
208         transaction = journal->j_running_transaction;
209         if (transaction && time_after_eq(jiffies, transaction->t_expires)) {
210                 journal->j_commit_request = transaction->t_tid;
211                 jbd_debug(1, "woke because of timeout\n");
212         }
213         goto loop;
214
215 end_loop:
216         write_unlock(&journal->j_state_lock);
217         del_timer_sync(&journal->j_commit_timer);
218         journal->j_task = NULL;
219         wake_up(&journal->j_wait_done_commit);
220         jbd_debug(1, "Journal thread exiting.\n");
221         return 0;
222 }
223
224 static int jbd2_journal_start_thread(journal_t *journal)
225 {
226         struct task_struct *t;
227
228         t = kthread_run(kjournald2, journal, "jbd2/%s",
229                         journal->j_devname);
230         if (IS_ERR(t))
231                 return PTR_ERR(t);
232
233         wait_event(journal->j_wait_done_commit, journal->j_task != NULL);
234         return 0;
235 }
236
237 static void journal_kill_thread(journal_t *journal)
238 {
239         write_lock(&journal->j_state_lock);
240         journal->j_flags |= JBD2_UNMOUNT;
241
242         while (journal->j_task) {
243                 wake_up(&journal->j_wait_commit);
244                 write_unlock(&journal->j_state_lock);
245                 wait_event(journal->j_wait_done_commit, journal->j_task == NULL);
246                 write_lock(&journal->j_state_lock);
247         }
248         write_unlock(&journal->j_state_lock);
249 }
250
251 /*
252  * jbd2_journal_write_metadata_buffer: write a metadata buffer to the journal.
253  *
254  * Writes a metadata buffer to a given disk block.  The actual IO is not
255  * performed but a new buffer_head is constructed which labels the data
256  * to be written with the correct destination disk block.
257  *
258  * Any magic-number escaping which needs to be done will cause a
259  * copy-out here.  If the buffer happens to start with the
260  * JBD2_MAGIC_NUMBER, then we can't write it to the log directly: the
261  * magic number is only written to the log for descripter blocks.  In
262  * this case, we copy the data and replace the first word with 0, and we
263  * return a result code which indicates that this buffer needs to be
264  * marked as an escaped buffer in the corresponding log descriptor
265  * block.  The missing word can then be restored when the block is read
266  * during recovery.
267  *
268  * If the source buffer has already been modified by a new transaction
269  * since we took the last commit snapshot, we use the frozen copy of
270  * that data for IO.  If we end up using the existing buffer_head's data
271  * for the write, then we *have* to lock the buffer to prevent anyone
272  * else from using and possibly modifying it while the IO is in
273  * progress.
274  *
275  * The function returns a pointer to the buffer_heads to be used for IO.
276  *
277  * We assume that the journal has already been locked in this function.
278  *
279  * Return value:
280  *  <0: Error
281  * >=0: Finished OK
282  *
283  * On success:
284  * Bit 0 set == escape performed on the data
285  * Bit 1 set == buffer copy-out performed (kfree the data after IO)
286  */
287
288 int jbd2_journal_write_metadata_buffer(transaction_t *transaction,
289                                   struct journal_head  *jh_in,
290                                   struct journal_head **jh_out,
291                                   unsigned long long blocknr)
292 {
293         int need_copy_out = 0;
294         int done_copy_out = 0;
295         int do_escape = 0;
296         char *mapped_data;
297         struct buffer_head *new_bh;
298         struct journal_head *new_jh;
299         struct page *new_page;
300         unsigned int new_offset;
301         struct buffer_head *bh_in = jh2bh(jh_in);
302         journal_t *journal = transaction->t_journal;
303
304         /*
305          * The buffer really shouldn't be locked: only the current committing
306          * transaction is allowed to write it, so nobody else is allowed
307          * to do any IO.
308          *
309          * akpm: except if we're journalling data, and write() output is
310          * also part of a shared mapping, and another thread has
311          * decided to launch a writepage() against this buffer.
312          */
313         J_ASSERT_BH(bh_in, buffer_jbddirty(bh_in));
314
315 retry_alloc:
316         new_bh = alloc_buffer_head(GFP_NOFS);
317         if (!new_bh) {
318                 /*
319                  * Failure is not an option, but __GFP_NOFAIL is going
320                  * away; so we retry ourselves here.
321                  */
322                 congestion_wait(BLK_RW_ASYNC, HZ/50);
323                 goto retry_alloc;
324         }
325
326         /* keep subsequent assertions sane */
327         new_bh->b_state = 0;
328         init_buffer(new_bh, NULL, NULL);
329         atomic_set(&new_bh->b_count, 1);
330         new_jh = jbd2_journal_add_journal_head(new_bh); /* This sleeps */
331
332         /*
333          * If a new transaction has already done a buffer copy-out, then
334          * we use that version of the data for the commit.
335          */
336         jbd_lock_bh_state(bh_in);
337 repeat:
338         if (jh_in->b_frozen_data) {
339                 done_copy_out = 1;
340                 new_page = virt_to_page(jh_in->b_frozen_data);
341                 new_offset = offset_in_page(jh_in->b_frozen_data);
342         } else {
343                 new_page = jh2bh(jh_in)->b_page;
344                 new_offset = offset_in_page(jh2bh(jh_in)->b_data);
345         }
346
347         mapped_data = kmap_atomic(new_page);
348         /*
349          * Fire data frozen trigger if data already wasn't frozen.  Do this
350          * before checking for escaping, as the trigger may modify the magic
351          * offset.  If a copy-out happens afterwards, it will have the correct
352          * data in the buffer.
353          */
354         if (!done_copy_out)
355                 jbd2_buffer_frozen_trigger(jh_in, mapped_data + new_offset,
356                                            jh_in->b_triggers);
357
358         /*
359          * Check for escaping
360          */
361         if (*((__be32 *)(mapped_data + new_offset)) ==
362                                 cpu_to_be32(JBD2_MAGIC_NUMBER)) {
363                 need_copy_out = 1;
364                 do_escape = 1;
365         }
366         kunmap_atomic(mapped_data);
367
368         /*
369          * Do we need to do a data copy?
370          */
371         if (need_copy_out && !done_copy_out) {
372                 char *tmp;
373
374                 jbd_unlock_bh_state(bh_in);
375                 tmp = jbd2_alloc(bh_in->b_size, GFP_NOFS);
376                 if (!tmp) {
377                         jbd2_journal_put_journal_head(new_jh);
378                         return -ENOMEM;
379                 }
380                 jbd_lock_bh_state(bh_in);
381                 if (jh_in->b_frozen_data) {
382                         jbd2_free(tmp, bh_in->b_size);
383                         goto repeat;
384                 }
385
386                 jh_in->b_frozen_data = tmp;
387                 mapped_data = kmap_atomic(new_page);
388                 memcpy(tmp, mapped_data + new_offset, jh2bh(jh_in)->b_size);
389                 kunmap_atomic(mapped_data);
390
391                 new_page = virt_to_page(tmp);
392                 new_offset = offset_in_page(tmp);
393                 done_copy_out = 1;
394
395                 /*
396                  * This isn't strictly necessary, as we're using frozen
397                  * data for the escaping, but it keeps consistency with
398                  * b_frozen_data usage.
399                  */
400                 jh_in->b_frozen_triggers = jh_in->b_triggers;
401         }
402
403         /*
404          * Did we need to do an escaping?  Now we've done all the
405          * copying, we can finally do so.
406          */
407         if (do_escape) {
408                 mapped_data = kmap_atomic(new_page);
409                 *((unsigned int *)(mapped_data + new_offset)) = 0;
410                 kunmap_atomic(mapped_data);
411         }
412
413         set_bh_page(new_bh, new_page, new_offset);
414         new_jh->b_transaction = NULL;
415         new_bh->b_size = jh2bh(jh_in)->b_size;
416         new_bh->b_bdev = transaction->t_journal->j_dev;
417         new_bh->b_blocknr = blocknr;
418         set_buffer_mapped(new_bh);
419         set_buffer_dirty(new_bh);
420
421         *jh_out = new_jh;
422
423         /*
424          * The to-be-written buffer needs to get moved to the io queue,
425          * and the original buffer whose contents we are shadowing or
426          * copying is moved to the transaction's shadow queue.
427          */
428         JBUFFER_TRACE(jh_in, "file as BJ_Shadow");
429         spin_lock(&journal->j_list_lock);
430         __jbd2_journal_file_buffer(jh_in, transaction, BJ_Shadow);
431         spin_unlock(&journal->j_list_lock);
432         jbd_unlock_bh_state(bh_in);
433
434         JBUFFER_TRACE(new_jh, "file as BJ_IO");
435         jbd2_journal_file_buffer(new_jh, transaction, BJ_IO);
436
437         return do_escape | (done_copy_out << 1);
438 }
439
440 /*
441  * Allocation code for the journal file.  Manage the space left in the
442  * journal, so that we can begin checkpointing when appropriate.
443  */
444
445 /*
446  * __jbd2_log_space_left: Return the number of free blocks left in the journal.
447  *
448  * Called with the journal already locked.
449  *
450  * Called under j_state_lock
451  */
452
453 int __jbd2_log_space_left(journal_t *journal)
454 {
455         int left = journal->j_free;
456
457         /* assert_spin_locked(&journal->j_state_lock); */
458
459         /*
460          * Be pessimistic here about the number of those free blocks which
461          * might be required for log descriptor control blocks.
462          */
463
464 #define MIN_LOG_RESERVED_BLOCKS 32 /* Allow for rounding errors */
465
466         left -= MIN_LOG_RESERVED_BLOCKS;
467
468         if (left <= 0)
469                 return 0;
470         left -= (left >> 3);
471         return left;
472 }
473
474 /*
475  * Called with j_state_lock locked for writing.
476  * Returns true if a transaction commit was started.
477  */
478 int __jbd2_log_start_commit(journal_t *journal, tid_t target)
479 {
480         /*
481          * The only transaction we can possibly wait upon is the
482          * currently running transaction (if it exists).  Otherwise,
483          * the target tid must be an old one.
484          */
485         if (journal->j_running_transaction &&
486             journal->j_running_transaction->t_tid == target) {
487                 /*
488                  * We want a new commit: OK, mark the request and wakeup the
489                  * commit thread.  We do _not_ do the commit ourselves.
490                  */
491
492                 journal->j_commit_request = target;
493                 jbd_debug(1, "JBD2: requesting commit %d/%d\n",
494                           journal->j_commit_request,
495                           journal->j_commit_sequence);
496                 wake_up(&journal->j_wait_commit);
497                 return 1;
498         } else if (!tid_geq(journal->j_commit_request, target))
499                 /* This should never happen, but if it does, preserve
500                    the evidence before kjournald goes into a loop and
501                    increments j_commit_sequence beyond all recognition. */
502                 WARN_ONCE(1, "JBD2: bad log_start_commit: %u %u %u %u\n",
503                           journal->j_commit_request,
504                           journal->j_commit_sequence,
505                           target, journal->j_running_transaction ? 
506                           journal->j_running_transaction->t_tid : 0);
507         return 0;
508 }
509
510 int jbd2_log_start_commit(journal_t *journal, tid_t tid)
511 {
512         int ret;
513
514         write_lock(&journal->j_state_lock);
515         ret = __jbd2_log_start_commit(journal, tid);
516         write_unlock(&journal->j_state_lock);
517         return ret;
518 }
519
520 /*
521  * Force and wait upon a commit if the calling process is not within
522  * transaction.  This is used for forcing out undo-protected data which contains
523  * bitmaps, when the fs is running out of space.
524  *
525  * We can only force the running transaction if we don't have an active handle;
526  * otherwise, we will deadlock.
527  *
528  * Returns true if a transaction was started.
529  */
530 int jbd2_journal_force_commit_nested(journal_t *journal)
531 {
532         transaction_t *transaction = NULL;
533         tid_t tid;
534         int need_to_start = 0;
535
536         read_lock(&journal->j_state_lock);
537         if (journal->j_running_transaction && !current->journal_info) {
538                 transaction = journal->j_running_transaction;
539                 if (!tid_geq(journal->j_commit_request, transaction->t_tid))
540                         need_to_start = 1;
541         } else if (journal->j_committing_transaction)
542                 transaction = journal->j_committing_transaction;
543
544         if (!transaction) {
545                 read_unlock(&journal->j_state_lock);
546                 return 0;       /* Nothing to retry */
547         }
548
549         tid = transaction->t_tid;
550         read_unlock(&journal->j_state_lock);
551         if (need_to_start)
552                 jbd2_log_start_commit(journal, tid);
553         jbd2_log_wait_commit(journal, tid);
554         return 1;
555 }
556
557 /*
558  * Start a commit of the current running transaction (if any).  Returns true
559  * if a transaction is going to be committed (or is currently already
560  * committing), and fills its tid in at *ptid
561  */
562 int jbd2_journal_start_commit(journal_t *journal, tid_t *ptid)
563 {
564         int ret = 0;
565
566         write_lock(&journal->j_state_lock);
567         if (journal->j_running_transaction) {
568                 tid_t tid = journal->j_running_transaction->t_tid;
569
570                 __jbd2_log_start_commit(journal, tid);
571                 /* There's a running transaction and we've just made sure
572                  * it's commit has been scheduled. */
573                 if (ptid)
574                         *ptid = tid;
575                 ret = 1;
576         } else if (journal->j_committing_transaction) {
577                 /*
578                  * If ext3_write_super() recently started a commit, then we
579                  * have to wait for completion of that transaction
580                  */
581                 if (ptid)
582                         *ptid = journal->j_committing_transaction->t_tid;
583                 ret = 1;
584         }
585         write_unlock(&journal->j_state_lock);
586         return ret;
587 }
588
589 /*
590  * Return 1 if a given transaction has not yet sent barrier request
591  * connected with a transaction commit. If 0 is returned, transaction
592  * may or may not have sent the barrier. Used to avoid sending barrier
593  * twice in common cases.
594  */
595 int jbd2_trans_will_send_data_barrier(journal_t *journal, tid_t tid)
596 {
597         int ret = 0;
598         transaction_t *commit_trans;
599
600         if (!(journal->j_flags & JBD2_BARRIER))
601                 return 0;
602         read_lock(&journal->j_state_lock);
603         /* Transaction already committed? */
604         if (tid_geq(journal->j_commit_sequence, tid))
605                 goto out;
606         commit_trans = journal->j_committing_transaction;
607         if (!commit_trans || commit_trans->t_tid != tid) {
608                 ret = 1;
609                 goto out;
610         }
611         /*
612          * Transaction is being committed and we already proceeded to
613          * submitting a flush to fs partition?
614          */
615         if (journal->j_fs_dev != journal->j_dev) {
616                 if (!commit_trans->t_need_data_flush ||
617                     commit_trans->t_state >= T_COMMIT_DFLUSH)
618                         goto out;
619         } else {
620                 if (commit_trans->t_state >= T_COMMIT_JFLUSH)
621                         goto out;
622         }
623         ret = 1;
624 out:
625         read_unlock(&journal->j_state_lock);
626         return ret;
627 }
628 EXPORT_SYMBOL(jbd2_trans_will_send_data_barrier);
629
630 /*
631  * Wait for a specified commit to complete.
632  * The caller may not hold the journal lock.
633  */
634 int jbd2_log_wait_commit(journal_t *journal, tid_t tid)
635 {
636         int err = 0;
637
638         read_lock(&journal->j_state_lock);
639 #ifdef CONFIG_JBD2_DEBUG
640         if (!tid_geq(journal->j_commit_request, tid)) {
641                 printk(KERN_EMERG
642                        "%s: error: j_commit_request=%d, tid=%d\n",
643                        __func__, journal->j_commit_request, tid);
644         }
645 #endif
646         while (tid_gt(tid, journal->j_commit_sequence)) {
647                 jbd_debug(1, "JBD2: want %d, j_commit_sequence=%d\n",
648                                   tid, journal->j_commit_sequence);
649                 wake_up(&journal->j_wait_commit);
650                 read_unlock(&journal->j_state_lock);
651                 wait_event(journal->j_wait_done_commit,
652                                 !tid_gt(tid, journal->j_commit_sequence));
653                 read_lock(&journal->j_state_lock);
654         }
655         read_unlock(&journal->j_state_lock);
656
657         if (unlikely(is_journal_aborted(journal))) {
658                 printk(KERN_EMERG "journal commit I/O error\n");
659                 err = -EIO;
660         }
661         return err;
662 }
663
664 /*
665  * Log buffer allocation routines:
666  */
667
668 int jbd2_journal_next_log_block(journal_t *journal, unsigned long long *retp)
669 {
670         unsigned long blocknr;
671
672         write_lock(&journal->j_state_lock);
673         J_ASSERT(journal->j_free > 1);
674
675         blocknr = journal->j_head;
676         journal->j_head++;
677         journal->j_free--;
678         if (journal->j_head == journal->j_last)
679                 journal->j_head = journal->j_first;
680         write_unlock(&journal->j_state_lock);
681         return jbd2_journal_bmap(journal, blocknr, retp);
682 }
683
684 /*
685  * Conversion of logical to physical block numbers for the journal
686  *
687  * On external journals the journal blocks are identity-mapped, so
688  * this is a no-op.  If needed, we can use j_blk_offset - everything is
689  * ready.
690  */
691 int jbd2_journal_bmap(journal_t *journal, unsigned long blocknr,
692                  unsigned long long *retp)
693 {
694         int err = 0;
695         unsigned long long ret;
696
697         if (journal->j_inode) {
698                 ret = bmap(journal->j_inode, blocknr);
699                 if (ret)
700                         *retp = ret;
701                 else {
702                         printk(KERN_ALERT "%s: journal block not found "
703                                         "at offset %lu on %s\n",
704                                __func__, blocknr, journal->j_devname);
705                         err = -EIO;
706                         __journal_abort_soft(journal, err);
707                 }
708         } else {
709                 *retp = blocknr; /* +journal->j_blk_offset */
710         }
711         return err;
712 }
713
714 /*
715  * We play buffer_head aliasing tricks to write data/metadata blocks to
716  * the journal without copying their contents, but for journal
717  * descriptor blocks we do need to generate bona fide buffers.
718  *
719  * After the caller of jbd2_journal_get_descriptor_buffer() has finished modifying
720  * the buffer's contents they really should run flush_dcache_page(bh->b_page).
721  * But we don't bother doing that, so there will be coherency problems with
722  * mmaps of blockdevs which hold live JBD-controlled filesystems.
723  */
724 struct journal_head *jbd2_journal_get_descriptor_buffer(journal_t *journal)
725 {
726         struct buffer_head *bh;
727         unsigned long long blocknr;
728         int err;
729
730         err = jbd2_journal_next_log_block(journal, &blocknr);
731
732         if (err)
733                 return NULL;
734
735         bh = __getblk(journal->j_dev, blocknr, journal->j_blocksize);
736         if (!bh)
737                 return NULL;
738         lock_buffer(bh);
739         memset(bh->b_data, 0, journal->j_blocksize);
740         set_buffer_uptodate(bh);
741         unlock_buffer(bh);
742         BUFFER_TRACE(bh, "return this buffer");
743         return jbd2_journal_add_journal_head(bh);
744 }
745
746 /*
747  * Return tid of the oldest transaction in the journal and block in the journal
748  * where the transaction starts.
749  *
750  * If the journal is now empty, return which will be the next transaction ID
751  * we will write and where will that transaction start.
752  *
753  * The return value is 0 if journal tail cannot be pushed any further, 1 if
754  * it can.
755  */
756 int jbd2_journal_get_log_tail(journal_t *journal, tid_t *tid,
757                               unsigned long *block)
758 {
759         transaction_t *transaction;
760         int ret;
761
762         read_lock(&journal->j_state_lock);
763         spin_lock(&journal->j_list_lock);
764         transaction = journal->j_checkpoint_transactions;
765         if (transaction) {
766                 *tid = transaction->t_tid;
767                 *block = transaction->t_log_start;
768         } else if ((transaction = journal->j_committing_transaction) != NULL) {
769                 *tid = transaction->t_tid;
770                 *block = transaction->t_log_start;
771         } else if ((transaction = journal->j_running_transaction) != NULL) {
772                 *tid = transaction->t_tid;
773                 *block = journal->j_head;
774         } else {
775                 *tid = journal->j_transaction_sequence;
776                 *block = journal->j_head;
777         }
778         ret = tid_gt(*tid, journal->j_tail_sequence);
779         spin_unlock(&journal->j_list_lock);
780         read_unlock(&journal->j_state_lock);
781
782         return ret;
783 }
784
785 /*
786  * Update information in journal structure and in on disk journal superblock
787  * about log tail. This function does not check whether information passed in
788  * really pushes log tail further. It's responsibility of the caller to make
789  * sure provided log tail information is valid (e.g. by holding
790  * j_checkpoint_mutex all the time between computing log tail and calling this
791  * function as is the case with jbd2_cleanup_journal_tail()).
792  *
793  * Requires j_checkpoint_mutex
794  */
795 void __jbd2_update_log_tail(journal_t *journal, tid_t tid, unsigned long block)
796 {
797         unsigned long freed;
798
799         BUG_ON(!mutex_is_locked(&journal->j_checkpoint_mutex));
800
801         /*
802          * We cannot afford for write to remain in drive's caches since as
803          * soon as we update j_tail, next transaction can start reusing journal
804          * space and if we lose sb update during power failure we'd replay
805          * old transaction with possibly newly overwritten data.
806          */
807         jbd2_journal_update_sb_log_tail(journal, tid, block, WRITE_FUA);
808         write_lock(&journal->j_state_lock);
809         freed = block - journal->j_tail;
810         if (block < journal->j_tail)
811                 freed += journal->j_last - journal->j_first;
812
813         trace_jbd2_update_log_tail(journal, tid, block, freed);
814         jbd_debug(1,
815                   "Cleaning journal tail from %d to %d (offset %lu), "
816                   "freeing %lu\n",
817                   journal->j_tail_sequence, tid, block, freed);
818
819         journal->j_free += freed;
820         journal->j_tail_sequence = tid;
821         journal->j_tail = block;
822         write_unlock(&journal->j_state_lock);
823 }
824
825 /*
826  * This is a variaon of __jbd2_update_log_tail which checks for validity of
827  * provided log tail and locks j_checkpoint_mutex. So it is safe against races
828  * with other threads updating log tail.
829  */
830 void jbd2_update_log_tail(journal_t *journal, tid_t tid, unsigned long block)
831 {
832         mutex_lock(&journal->j_checkpoint_mutex);
833         if (tid_gt(tid, journal->j_tail_sequence))
834                 __jbd2_update_log_tail(journal, tid, block);
835         mutex_unlock(&journal->j_checkpoint_mutex);
836 }
837
838 struct jbd2_stats_proc_session {
839         journal_t *journal;
840         struct transaction_stats_s *stats;
841         int start;
842         int max;
843 };
844
845 static void *jbd2_seq_info_start(struct seq_file *seq, loff_t *pos)
846 {
847         return *pos ? NULL : SEQ_START_TOKEN;
848 }
849
850 static void *jbd2_seq_info_next(struct seq_file *seq, void *v, loff_t *pos)
851 {
852         return NULL;
853 }
854
855 static int jbd2_seq_info_show(struct seq_file *seq, void *v)
856 {
857         struct jbd2_stats_proc_session *s = seq->private;
858
859         if (v != SEQ_START_TOKEN)
860                 return 0;
861         seq_printf(seq, "%lu transaction, each up to %u blocks\n",
862                         s->stats->ts_tid,
863                         s->journal->j_max_transaction_buffers);
864         if (s->stats->ts_tid == 0)
865                 return 0;
866         seq_printf(seq, "average: \n  %ums waiting for transaction\n",
867             jiffies_to_msecs(s->stats->run.rs_wait / s->stats->ts_tid));
868         seq_printf(seq, "  %ums running transaction\n",
869             jiffies_to_msecs(s->stats->run.rs_running / s->stats->ts_tid));
870         seq_printf(seq, "  %ums transaction was being locked\n",
871             jiffies_to_msecs(s->stats->run.rs_locked / s->stats->ts_tid));
872         seq_printf(seq, "  %ums flushing data (in ordered mode)\n",
873             jiffies_to_msecs(s->stats->run.rs_flushing / s->stats->ts_tid));
874         seq_printf(seq, "  %ums logging transaction\n",
875             jiffies_to_msecs(s->stats->run.rs_logging / s->stats->ts_tid));
876         seq_printf(seq, "  %lluus average transaction commit time\n",
877                    div_u64(s->journal->j_average_commit_time, 1000));
878         seq_printf(seq, "  %lu handles per transaction\n",
879             s->stats->run.rs_handle_count / s->stats->ts_tid);
880         seq_printf(seq, "  %lu blocks per transaction\n",
881             s->stats->run.rs_blocks / s->stats->ts_tid);
882         seq_printf(seq, "  %lu logged blocks per transaction\n",
883             s->stats->run.rs_blocks_logged / s->stats->ts_tid);
884         return 0;
885 }
886
887 static void jbd2_seq_info_stop(struct seq_file *seq, void *v)
888 {
889 }
890
891 static const struct seq_operations jbd2_seq_info_ops = {
892         .start  = jbd2_seq_info_start,
893         .next   = jbd2_seq_info_next,
894         .stop   = jbd2_seq_info_stop,
895         .show   = jbd2_seq_info_show,
896 };
897
898 static int jbd2_seq_info_open(struct inode *inode, struct file *file)
899 {
900         journal_t *journal = PDE(inode)->data;
901         struct jbd2_stats_proc_session *s;
902         int rc, size;
903
904         s = kmalloc(sizeof(*s), GFP_KERNEL);
905         if (s == NULL)
906                 return -ENOMEM;
907         size = sizeof(struct transaction_stats_s);
908         s->stats = kmalloc(size, GFP_KERNEL);
909         if (s->stats == NULL) {
910                 kfree(s);
911                 return -ENOMEM;
912         }
913         spin_lock(&journal->j_history_lock);
914         memcpy(s->stats, &journal->j_stats, size);
915         s->journal = journal;
916         spin_unlock(&journal->j_history_lock);
917
918         rc = seq_open(file, &jbd2_seq_info_ops);
919         if (rc == 0) {
920                 struct seq_file *m = file->private_data;
921                 m->private = s;
922         } else {
923                 kfree(s->stats);
924                 kfree(s);
925         }
926         return rc;
927
928 }
929
930 static int jbd2_seq_info_release(struct inode *inode, struct file *file)
931 {
932         struct seq_file *seq = file->private_data;
933         struct jbd2_stats_proc_session *s = seq->private;
934         kfree(s->stats);
935         kfree(s);
936         return seq_release(inode, file);
937 }
938
939 static const struct file_operations jbd2_seq_info_fops = {
940         .owner          = THIS_MODULE,
941         .open           = jbd2_seq_info_open,
942         .read           = seq_read,
943         .llseek         = seq_lseek,
944         .release        = jbd2_seq_info_release,
945 };
946
947 static struct proc_dir_entry *proc_jbd2_stats;
948
949 static void jbd2_stats_proc_init(journal_t *journal)
950 {
951         journal->j_proc_entry = proc_mkdir(journal->j_devname, proc_jbd2_stats);
952         if (journal->j_proc_entry) {
953                 proc_create_data("info", S_IRUGO, journal->j_proc_entry,
954                                  &jbd2_seq_info_fops, journal);
955         }
956 }
957
958 static void jbd2_stats_proc_exit(journal_t *journal)
959 {
960         remove_proc_entry("info", journal->j_proc_entry);
961         remove_proc_entry(journal->j_devname, proc_jbd2_stats);
962 }
963
964 /*
965  * Management for journal control blocks: functions to create and
966  * destroy journal_t structures, and to initialise and read existing
967  * journal blocks from disk.  */
968
969 /* First: create and setup a journal_t object in memory.  We initialise
970  * very few fields yet: that has to wait until we have created the
971  * journal structures from from scratch, or loaded them from disk. */
972
973 static journal_t * journal_init_common (void)
974 {
975         journal_t *journal;
976         int err;
977
978         journal = kzalloc(sizeof(*journal), GFP_KERNEL);
979         if (!journal)
980                 return NULL;
981
982         init_waitqueue_head(&journal->j_wait_transaction_locked);
983         init_waitqueue_head(&journal->j_wait_logspace);
984         init_waitqueue_head(&journal->j_wait_done_commit);
985         init_waitqueue_head(&journal->j_wait_checkpoint);
986         init_waitqueue_head(&journal->j_wait_commit);
987         init_waitqueue_head(&journal->j_wait_updates);
988         mutex_init(&journal->j_barrier);
989         mutex_init(&journal->j_checkpoint_mutex);
990         spin_lock_init(&journal->j_revoke_lock);
991         spin_lock_init(&journal->j_list_lock);
992         rwlock_init(&journal->j_state_lock);
993
994         journal->j_commit_interval = (HZ * JBD2_DEFAULT_MAX_COMMIT_AGE);
995         journal->j_min_batch_time = 0;
996         journal->j_max_batch_time = 15000; /* 15ms */
997
998         /* The journal is marked for error until we succeed with recovery! */
999         journal->j_flags = JBD2_ABORT;
1000
1001         /* Set up a default-sized revoke table for the new mount. */
1002         err = jbd2_journal_init_revoke(journal, JOURNAL_REVOKE_DEFAULT_HASH);
1003         if (err) {
1004                 kfree(journal);
1005                 return NULL;
1006         }
1007
1008         spin_lock_init(&journal->j_history_lock);
1009
1010         return journal;
1011 }
1012
1013 /* jbd2_journal_init_dev and jbd2_journal_init_inode:
1014  *
1015  * Create a journal structure assigned some fixed set of disk blocks to
1016  * the journal.  We don't actually touch those disk blocks yet, but we
1017  * need to set up all of the mapping information to tell the journaling
1018  * system where the journal blocks are.
1019  *
1020  */
1021
1022 /**
1023  *  journal_t * jbd2_journal_init_dev() - creates and initialises a journal structure
1024  *  @bdev: Block device on which to create the journal
1025  *  @fs_dev: Device which hold journalled filesystem for this journal.
1026  *  @start: Block nr Start of journal.
1027  *  @len:  Length of the journal in blocks.
1028  *  @blocksize: blocksize of journalling device
1029  *
1030  *  Returns: a newly created journal_t *
1031  *
1032  *  jbd2_journal_init_dev creates a journal which maps a fixed contiguous
1033  *  range of blocks on an arbitrary block device.
1034  *
1035  */
1036 journal_t * jbd2_journal_init_dev(struct block_device *bdev,
1037                         struct block_device *fs_dev,
1038                         unsigned long long start, int len, int blocksize)
1039 {
1040         journal_t *journal = journal_init_common();
1041         struct buffer_head *bh;
1042         char *p;
1043         int n;
1044
1045         if (!journal)
1046                 return NULL;
1047
1048         /* journal descriptor can store up to n blocks -bzzz */
1049         journal->j_blocksize = blocksize;
1050         journal->j_dev = bdev;
1051         journal->j_fs_dev = fs_dev;
1052         journal->j_blk_offset = start;
1053         journal->j_maxlen = len;
1054         bdevname(journal->j_dev, journal->j_devname);
1055         p = journal->j_devname;
1056         while ((p = strchr(p, '/')))
1057                 *p = '!';
1058         jbd2_stats_proc_init(journal);
1059         n = journal->j_blocksize / sizeof(journal_block_tag_t);
1060         journal->j_wbufsize = n;
1061         journal->j_wbuf = kmalloc(n * sizeof(struct buffer_head*), GFP_KERNEL);
1062         if (!journal->j_wbuf) {
1063                 printk(KERN_ERR "%s: Can't allocate bhs for commit thread\n",
1064                         __func__);
1065                 goto out_err;
1066         }
1067
1068         bh = __getblk(journal->j_dev, start, journal->j_blocksize);
1069         if (!bh) {
1070                 printk(KERN_ERR
1071                        "%s: Cannot get buffer for journal superblock\n",
1072                        __func__);
1073                 goto out_err;
1074         }
1075         journal->j_sb_buffer = bh;
1076         journal->j_superblock = (journal_superblock_t *)bh->b_data;
1077
1078         return journal;
1079 out_err:
1080         kfree(journal->j_wbuf);
1081         jbd2_stats_proc_exit(journal);
1082         kfree(journal);
1083         return NULL;
1084 }
1085
1086 /**
1087  *  journal_t * jbd2_journal_init_inode () - creates a journal which maps to a inode.
1088  *  @inode: An inode to create the journal in
1089  *
1090  * jbd2_journal_init_inode creates a journal which maps an on-disk inode as
1091  * the journal.  The inode must exist already, must support bmap() and
1092  * must have all data blocks preallocated.
1093  */
1094 journal_t * jbd2_journal_init_inode (struct inode *inode)
1095 {
1096         struct buffer_head *bh;
1097         journal_t *journal = journal_init_common();
1098         char *p;
1099         int err;
1100         int n;
1101         unsigned long long blocknr;
1102
1103         if (!journal)
1104                 return NULL;
1105
1106         journal->j_dev = journal->j_fs_dev = inode->i_sb->s_bdev;
1107         journal->j_inode = inode;
1108         bdevname(journal->j_dev, journal->j_devname);
1109         p = journal->j_devname;
1110         while ((p = strchr(p, '/')))
1111                 *p = '!';
1112         p = journal->j_devname + strlen(journal->j_devname);
1113         sprintf(p, "-%lu", journal->j_inode->i_ino);
1114         jbd_debug(1,
1115                   "journal %p: inode %s/%ld, size %Ld, bits %d, blksize %ld\n",
1116                   journal, inode->i_sb->s_id, inode->i_ino,
1117                   (long long) inode->i_size,
1118                   inode->i_sb->s_blocksize_bits, inode->i_sb->s_blocksize);
1119
1120         journal->j_maxlen = inode->i_size >> inode->i_sb->s_blocksize_bits;
1121         journal->j_blocksize = inode->i_sb->s_blocksize;
1122         jbd2_stats_proc_init(journal);
1123
1124         /* journal descriptor can store up to n blocks -bzzz */
1125         n = journal->j_blocksize / sizeof(journal_block_tag_t);
1126         journal->j_wbufsize = n;
1127         journal->j_wbuf = kmalloc(n * sizeof(struct buffer_head*), GFP_KERNEL);
1128         if (!journal->j_wbuf) {
1129                 printk(KERN_ERR "%s: Can't allocate bhs for commit thread\n",
1130                         __func__);
1131                 goto out_err;
1132         }
1133
1134         err = jbd2_journal_bmap(journal, 0, &blocknr);
1135         /* If that failed, give up */
1136         if (err) {
1137                 printk(KERN_ERR "%s: Cannot locate journal superblock\n",
1138                        __func__);
1139                 goto out_err;
1140         }
1141
1142         bh = __getblk(journal->j_dev, blocknr, journal->j_blocksize);
1143         if (!bh) {
1144                 printk(KERN_ERR
1145                        "%s: Cannot get buffer for journal superblock\n",
1146                        __func__);
1147                 goto out_err;
1148         }
1149         journal->j_sb_buffer = bh;
1150         journal->j_superblock = (journal_superblock_t *)bh->b_data;
1151
1152         return journal;
1153 out_err:
1154         kfree(journal->j_wbuf);
1155         jbd2_stats_proc_exit(journal);
1156         kfree(journal);
1157         return NULL;
1158 }
1159
1160 /*
1161  * If the journal init or create aborts, we need to mark the journal
1162  * superblock as being NULL to prevent the journal destroy from writing
1163  * back a bogus superblock.
1164  */
1165 static void journal_fail_superblock (journal_t *journal)
1166 {
1167         struct buffer_head *bh = journal->j_sb_buffer;
1168         brelse(bh);
1169         journal->j_sb_buffer = NULL;
1170 }
1171
1172 /*
1173  * Given a journal_t structure, initialise the various fields for
1174  * startup of a new journaling session.  We use this both when creating
1175  * a journal, and after recovering an old journal to reset it for
1176  * subsequent use.
1177  */
1178
1179 static int journal_reset(journal_t *journal)
1180 {
1181         journal_superblock_t *sb = journal->j_superblock;
1182         unsigned long long first, last;
1183
1184         first = be32_to_cpu(sb->s_first);
1185         last = be32_to_cpu(sb->s_maxlen);
1186         if (first + JBD2_MIN_JOURNAL_BLOCKS > last + 1) {
1187                 printk(KERN_ERR "JBD2: Journal too short (blocks %llu-%llu).\n",
1188                        first, last);
1189                 journal_fail_superblock(journal);
1190                 return -EINVAL;
1191         }
1192
1193         journal->j_first = first;
1194         journal->j_last = last;
1195
1196         journal->j_head = first;
1197         journal->j_tail = first;
1198         journal->j_free = last - first;
1199
1200         journal->j_tail_sequence = journal->j_transaction_sequence;
1201         journal->j_commit_sequence = journal->j_transaction_sequence - 1;
1202         journal->j_commit_request = journal->j_commit_sequence;
1203
1204         journal->j_max_transaction_buffers = journal->j_maxlen / 4;
1205
1206         /*
1207          * As a special case, if the on-disk copy is already marked as needing
1208          * no recovery (s_start == 0), then we can safely defer the superblock
1209          * update until the next commit by setting JBD2_FLUSHED.  This avoids
1210          * attempting a write to a potential-readonly device.
1211          */
1212         if (sb->s_start == 0) {
1213                 jbd_debug(1, "JBD2: Skipping superblock update on recovered sb "
1214                         "(start %ld, seq %d, errno %d)\n",
1215                         journal->j_tail, journal->j_tail_sequence,
1216                         journal->j_errno);
1217                 journal->j_flags |= JBD2_FLUSHED;
1218         } else {
1219                 /* Lock here to make assertions happy... */
1220                 mutex_lock(&journal->j_checkpoint_mutex);
1221                 /*
1222                  * Update log tail information. We use WRITE_FUA since new
1223                  * transaction will start reusing journal space and so we
1224                  * must make sure information about current log tail is on
1225                  * disk before that.
1226                  */
1227                 jbd2_journal_update_sb_log_tail(journal,
1228                                                 journal->j_tail_sequence,
1229                                                 journal->j_tail,
1230                                                 WRITE_FUA);
1231                 mutex_unlock(&journal->j_checkpoint_mutex);
1232         }
1233         return jbd2_journal_start_thread(journal);
1234 }
1235
1236 static void jbd2_write_superblock(journal_t *journal, int write_op)
1237 {
1238         struct buffer_head *bh = journal->j_sb_buffer;
1239         int ret;
1240
1241         trace_jbd2_write_superblock(journal, write_op);
1242         if (!(journal->j_flags & JBD2_BARRIER))
1243                 write_op &= ~(REQ_FUA | REQ_FLUSH);
1244         lock_buffer(bh);
1245         if (buffer_write_io_error(bh)) {
1246                 /*
1247                  * Oh, dear.  A previous attempt to write the journal
1248                  * superblock failed.  This could happen because the
1249                  * USB device was yanked out.  Or it could happen to
1250                  * be a transient write error and maybe the block will
1251                  * be remapped.  Nothing we can do but to retry the
1252                  * write and hope for the best.
1253                  */
1254                 printk(KERN_ERR "JBD2: previous I/O error detected "
1255                        "for journal superblock update for %s.\n",
1256                        journal->j_devname);
1257                 clear_buffer_write_io_error(bh);
1258                 set_buffer_uptodate(bh);
1259         }
1260         get_bh(bh);
1261         bh->b_end_io = end_buffer_write_sync;
1262         ret = submit_bh(write_op, bh);
1263         wait_on_buffer(bh);
1264         if (buffer_write_io_error(bh)) {
1265                 clear_buffer_write_io_error(bh);
1266                 set_buffer_uptodate(bh);
1267                 ret = -EIO;
1268         }
1269         if (ret) {
1270                 printk(KERN_ERR "JBD2: Error %d detected when updating "
1271                        "journal superblock for %s.\n", ret,
1272                        journal->j_devname);
1273         }
1274 }
1275
1276 /**
1277  * jbd2_journal_update_sb_log_tail() - Update log tail in journal sb on disk.
1278  * @journal: The journal to update.
1279  * @tail_tid: TID of the new transaction at the tail of the log
1280  * @tail_block: The first block of the transaction at the tail of the log
1281  * @write_op: With which operation should we write the journal sb
1282  *
1283  * Update a journal's superblock information about log tail and write it to
1284  * disk, waiting for the IO to complete.
1285  */
1286 void jbd2_journal_update_sb_log_tail(journal_t *journal, tid_t tail_tid,
1287                                      unsigned long tail_block, int write_op)
1288 {
1289         journal_superblock_t *sb = journal->j_superblock;
1290
1291         BUG_ON(!mutex_is_locked(&journal->j_checkpoint_mutex));
1292         jbd_debug(1, "JBD2: updating superblock (start %lu, seq %u)\n",
1293                   tail_block, tail_tid);
1294
1295         sb->s_sequence = cpu_to_be32(tail_tid);
1296         sb->s_start    = cpu_to_be32(tail_block);
1297
1298         jbd2_write_superblock(journal, write_op);
1299
1300         /* Log is no longer empty */
1301         write_lock(&journal->j_state_lock);
1302         WARN_ON(!sb->s_sequence);
1303         journal->j_flags &= ~JBD2_FLUSHED;
1304         write_unlock(&journal->j_state_lock);
1305 }
1306
1307 /**
1308  * jbd2_mark_journal_empty() - Mark on disk journal as empty.
1309  * @journal: The journal to update.
1310  *
1311  * Update a journal's dynamic superblock fields to show that journal is empty.
1312  * Write updated superblock to disk waiting for IO to complete.
1313  */
1314 static void jbd2_mark_journal_empty(journal_t *journal)
1315 {
1316         journal_superblock_t *sb = journal->j_superblock;
1317
1318         BUG_ON(!mutex_is_locked(&journal->j_checkpoint_mutex));
1319         read_lock(&journal->j_state_lock);
1320         jbd_debug(1, "JBD2: Marking journal as empty (seq %d)\n",
1321                   journal->j_tail_sequence);
1322
1323         sb->s_sequence = cpu_to_be32(journal->j_tail_sequence);
1324         sb->s_start    = cpu_to_be32(0);
1325         read_unlock(&journal->j_state_lock);
1326
1327         jbd2_write_superblock(journal, WRITE_FUA);
1328
1329         /* Log is no longer empty */
1330         write_lock(&journal->j_state_lock);
1331         journal->j_flags |= JBD2_FLUSHED;
1332         write_unlock(&journal->j_state_lock);
1333 }
1334
1335
1336 /**
1337  * jbd2_journal_update_sb_errno() - Update error in the journal.
1338  * @journal: The journal to update.
1339  *
1340  * Update a journal's errno.  Write updated superblock to disk waiting for IO
1341  * to complete.
1342  */
1343 static void jbd2_journal_update_sb_errno(journal_t *journal)
1344 {
1345         journal_superblock_t *sb = journal->j_superblock;
1346
1347         read_lock(&journal->j_state_lock);
1348         jbd_debug(1, "JBD2: updating superblock error (errno %d)\n",
1349                   journal->j_errno);
1350         sb->s_errno    = cpu_to_be32(journal->j_errno);
1351         read_unlock(&journal->j_state_lock);
1352
1353         jbd2_write_superblock(journal, WRITE_SYNC);
1354 }
1355
1356 /*
1357  * Read the superblock for a given journal, performing initial
1358  * validation of the format.
1359  */
1360 static int journal_get_superblock(journal_t *journal)
1361 {
1362         struct buffer_head *bh;
1363         journal_superblock_t *sb;
1364         int err = -EIO;
1365
1366         bh = journal->j_sb_buffer;
1367
1368         J_ASSERT(bh != NULL);
1369         if (!buffer_uptodate(bh)) {
1370                 ll_rw_block(READ, 1, &bh);
1371                 wait_on_buffer(bh);
1372                 if (!buffer_uptodate(bh)) {
1373                         printk(KERN_ERR
1374                                 "JBD2: IO error reading journal superblock\n");
1375                         goto out;
1376                 }
1377         }
1378
1379         sb = journal->j_superblock;
1380
1381         err = -EINVAL;
1382
1383         if (sb->s_header.h_magic != cpu_to_be32(JBD2_MAGIC_NUMBER) ||
1384             sb->s_blocksize != cpu_to_be32(journal->j_blocksize)) {
1385                 printk(KERN_WARNING "JBD2: no valid journal superblock found\n");
1386                 goto out;
1387         }
1388
1389         switch(be32_to_cpu(sb->s_header.h_blocktype)) {
1390         case JBD2_SUPERBLOCK_V1:
1391                 journal->j_format_version = 1;
1392                 break;
1393         case JBD2_SUPERBLOCK_V2:
1394                 journal->j_format_version = 2;
1395                 break;
1396         default:
1397                 printk(KERN_WARNING "JBD2: unrecognised superblock format ID\n");
1398                 goto out;
1399         }
1400
1401         if (be32_to_cpu(sb->s_maxlen) < journal->j_maxlen)
1402                 journal->j_maxlen = be32_to_cpu(sb->s_maxlen);
1403         else if (be32_to_cpu(sb->s_maxlen) > journal->j_maxlen) {
1404                 printk(KERN_WARNING "JBD2: journal file too short\n");
1405                 goto out;
1406         }
1407
1408         if (be32_to_cpu(sb->s_first) == 0 ||
1409             be32_to_cpu(sb->s_first) >= journal->j_maxlen) {
1410                 printk(KERN_WARNING
1411                         "JBD2: Invalid start block of journal: %u\n",
1412                         be32_to_cpu(sb->s_first));
1413                 goto out;
1414         }
1415
1416         return 0;
1417
1418 out:
1419         journal_fail_superblock(journal);
1420         return err;
1421 }
1422
1423 /*
1424  * Load the on-disk journal superblock and read the key fields into the
1425  * journal_t.
1426  */
1427
1428 static int load_superblock(journal_t *journal)
1429 {
1430         int err;
1431         journal_superblock_t *sb;
1432
1433         err = journal_get_superblock(journal);
1434         if (err)
1435                 return err;
1436
1437         sb = journal->j_superblock;
1438
1439         journal->j_tail_sequence = be32_to_cpu(sb->s_sequence);
1440         journal->j_tail = be32_to_cpu(sb->s_start);
1441         journal->j_first = be32_to_cpu(sb->s_first);
1442         journal->j_last = be32_to_cpu(sb->s_maxlen);
1443         journal->j_errno = be32_to_cpu(sb->s_errno);
1444
1445         return 0;
1446 }
1447
1448
1449 /**
1450  * int jbd2_journal_load() - Read journal from disk.
1451  * @journal: Journal to act on.
1452  *
1453  * Given a journal_t structure which tells us which disk blocks contain
1454  * a journal, read the journal from disk to initialise the in-memory
1455  * structures.
1456  */
1457 int jbd2_journal_load(journal_t *journal)
1458 {
1459         int err;
1460         journal_superblock_t *sb;
1461
1462         err = load_superblock(journal);
1463         if (err)
1464                 return err;
1465
1466         sb = journal->j_superblock;
1467         /* If this is a V2 superblock, then we have to check the
1468          * features flags on it. */
1469
1470         if (journal->j_format_version >= 2) {
1471                 if ((sb->s_feature_ro_compat &
1472                      ~cpu_to_be32(JBD2_KNOWN_ROCOMPAT_FEATURES)) ||
1473                     (sb->s_feature_incompat &
1474                      ~cpu_to_be32(JBD2_KNOWN_INCOMPAT_FEATURES))) {
1475                         printk(KERN_WARNING
1476                                 "JBD2: Unrecognised features on journal\n");
1477                         return -EINVAL;
1478                 }
1479         }
1480
1481         /*
1482          * Create a slab for this blocksize
1483          */
1484         err = jbd2_journal_create_slab(be32_to_cpu(sb->s_blocksize));
1485         if (err)
1486                 return err;
1487
1488         /* Let the recovery code check whether it needs to recover any
1489          * data from the journal. */
1490         if (jbd2_journal_recover(journal))
1491                 goto recovery_error;
1492
1493         if (journal->j_failed_commit) {
1494                 printk(KERN_ERR "JBD2: journal transaction %u on %s "
1495                        "is corrupt.\n", journal->j_failed_commit,
1496                        journal->j_devname);
1497                 return -EIO;
1498         }
1499
1500         /* OK, we've finished with the dynamic journal bits:
1501          * reinitialise the dynamic contents of the superblock in memory
1502          * and reset them on disk. */
1503         if (journal_reset(journal))
1504                 goto recovery_error;
1505
1506         journal->j_flags &= ~JBD2_ABORT;
1507         journal->j_flags |= JBD2_LOADED;
1508         return 0;
1509
1510 recovery_error:
1511         printk(KERN_WARNING "JBD2: recovery failed\n");
1512         return -EIO;
1513 }
1514
1515 /**
1516  * void jbd2_journal_destroy() - Release a journal_t structure.
1517  * @journal: Journal to act on.
1518  *
1519  * Release a journal_t structure once it is no longer in use by the
1520  * journaled object.
1521  * Return <0 if we couldn't clean up the journal.
1522  */
1523 int jbd2_journal_destroy(journal_t *journal)
1524 {
1525         int err = 0;
1526
1527         /* Wait for the commit thread to wake up and die. */
1528         journal_kill_thread(journal);
1529
1530         /* Force a final log commit */
1531         if (journal->j_running_transaction)
1532                 jbd2_journal_commit_transaction(journal);
1533
1534         /* Force any old transactions to disk */
1535
1536         /* Totally anal locking here... */
1537         spin_lock(&journal->j_list_lock);
1538         while (journal->j_checkpoint_transactions != NULL) {
1539                 spin_unlock(&journal->j_list_lock);
1540                 mutex_lock(&journal->j_checkpoint_mutex);
1541                 jbd2_log_do_checkpoint(journal);
1542                 mutex_unlock(&journal->j_checkpoint_mutex);
1543                 spin_lock(&journal->j_list_lock);
1544         }
1545
1546         J_ASSERT(journal->j_running_transaction == NULL);
1547         J_ASSERT(journal->j_committing_transaction == NULL);
1548         J_ASSERT(journal->j_checkpoint_transactions == NULL);
1549         spin_unlock(&journal->j_list_lock);
1550
1551         if (journal->j_sb_buffer) {
1552                 if (!is_journal_aborted(journal)) {
1553                         mutex_lock(&journal->j_checkpoint_mutex);
1554                         jbd2_mark_journal_empty(journal);
1555                         mutex_unlock(&journal->j_checkpoint_mutex);
1556                 } else
1557                         err = -EIO;
1558                 brelse(journal->j_sb_buffer);
1559         }
1560
1561         if (journal->j_proc_entry)
1562                 jbd2_stats_proc_exit(journal);
1563         if (journal->j_inode)
1564                 iput(journal->j_inode);
1565         if (journal->j_revoke)
1566                 jbd2_journal_destroy_revoke(journal);
1567         kfree(journal->j_wbuf);
1568         kfree(journal);
1569
1570         return err;
1571 }
1572
1573
1574 /**
1575  *int jbd2_journal_check_used_features () - Check if features specified are used.
1576  * @journal: Journal to check.
1577  * @compat: bitmask of compatible features
1578  * @ro: bitmask of features that force read-only mount
1579  * @incompat: bitmask of incompatible features
1580  *
1581  * Check whether the journal uses all of a given set of
1582  * features.  Return true (non-zero) if it does.
1583  **/
1584
1585 int jbd2_journal_check_used_features (journal_t *journal, unsigned long compat,
1586                                  unsigned long ro, unsigned long incompat)
1587 {
1588         journal_superblock_t *sb;
1589
1590         if (!compat && !ro && !incompat)
1591                 return 1;
1592         /* Load journal superblock if it is not loaded yet. */
1593         if (journal->j_format_version == 0 &&
1594             journal_get_superblock(journal) != 0)
1595                 return 0;
1596         if (journal->j_format_version == 1)
1597                 return 0;
1598
1599         sb = journal->j_superblock;
1600
1601         if (((be32_to_cpu(sb->s_feature_compat) & compat) == compat) &&
1602             ((be32_to_cpu(sb->s_feature_ro_compat) & ro) == ro) &&
1603             ((be32_to_cpu(sb->s_feature_incompat) & incompat) == incompat))
1604                 return 1;
1605
1606         return 0;
1607 }
1608
1609 /**
1610  * int jbd2_journal_check_available_features() - Check feature set in journalling layer
1611  * @journal: Journal to check.
1612  * @compat: bitmask of compatible features
1613  * @ro: bitmask of features that force read-only mount
1614  * @incompat: bitmask of incompatible features
1615  *
1616  * Check whether the journaling code supports the use of
1617  * all of a given set of features on this journal.  Return true
1618  * (non-zero) if it can. */
1619
1620 int jbd2_journal_check_available_features (journal_t *journal, unsigned long compat,
1621                                       unsigned long ro, unsigned long incompat)
1622 {
1623         if (!compat && !ro && !incompat)
1624                 return 1;
1625
1626         /* We can support any known requested features iff the
1627          * superblock is in version 2.  Otherwise we fail to support any
1628          * extended sb features. */
1629
1630         if (journal->j_format_version != 2)
1631                 return 0;
1632
1633         if ((compat   & JBD2_KNOWN_COMPAT_FEATURES) == compat &&
1634             (ro       & JBD2_KNOWN_ROCOMPAT_FEATURES) == ro &&
1635             (incompat & JBD2_KNOWN_INCOMPAT_FEATURES) == incompat)
1636                 return 1;
1637
1638         return 0;
1639 }
1640
1641 /**
1642  * int jbd2_journal_set_features () - Mark a given journal feature in the superblock
1643  * @journal: Journal to act on.
1644  * @compat: bitmask of compatible features
1645  * @ro: bitmask of features that force read-only mount
1646  * @incompat: bitmask of incompatible features
1647  *
1648  * Mark a given journal feature as present on the
1649  * superblock.  Returns true if the requested features could be set.
1650  *
1651  */
1652
1653 int jbd2_journal_set_features (journal_t *journal, unsigned long compat,
1654                           unsigned long ro, unsigned long incompat)
1655 {
1656         journal_superblock_t *sb;
1657
1658         if (jbd2_journal_check_used_features(journal, compat, ro, incompat))
1659                 return 1;
1660
1661         if (!jbd2_journal_check_available_features(journal, compat, ro, incompat))
1662                 return 0;
1663
1664         jbd_debug(1, "Setting new features 0x%lx/0x%lx/0x%lx\n",
1665                   compat, ro, incompat);
1666
1667         sb = journal->j_superblock;
1668
1669         sb->s_feature_compat    |= cpu_to_be32(compat);
1670         sb->s_feature_ro_compat |= cpu_to_be32(ro);
1671         sb->s_feature_incompat  |= cpu_to_be32(incompat);
1672
1673         return 1;
1674 }
1675
1676 /*
1677  * jbd2_journal_clear_features () - Clear a given journal feature in the
1678  *                                  superblock
1679  * @journal: Journal to act on.
1680  * @compat: bitmask of compatible features
1681  * @ro: bitmask of features that force read-only mount
1682  * @incompat: bitmask of incompatible features
1683  *
1684  * Clear a given journal feature as present on the
1685  * superblock.
1686  */
1687 void jbd2_journal_clear_features(journal_t *journal, unsigned long compat,
1688                                 unsigned long ro, unsigned long incompat)
1689 {
1690         journal_superblock_t *sb;
1691
1692         jbd_debug(1, "Clear features 0x%lx/0x%lx/0x%lx\n",
1693                   compat, ro, incompat);
1694
1695         sb = journal->j_superblock;
1696
1697         sb->s_feature_compat    &= ~cpu_to_be32(compat);
1698         sb->s_feature_ro_compat &= ~cpu_to_be32(ro);
1699         sb->s_feature_incompat  &= ~cpu_to_be32(incompat);
1700 }
1701 EXPORT_SYMBOL(jbd2_journal_clear_features);
1702
1703 /**
1704  * int jbd2_journal_flush () - Flush journal
1705  * @journal: Journal to act on.
1706  *
1707  * Flush all data for a given journal to disk and empty the journal.
1708  * Filesystems can use this when remounting readonly to ensure that
1709  * recovery does not need to happen on remount.
1710  */
1711
1712 int jbd2_journal_flush(journal_t *journal)
1713 {
1714         int err = 0;
1715         transaction_t *transaction = NULL;
1716
1717         write_lock(&journal->j_state_lock);
1718
1719         /* Force everything buffered to the log... */
1720         if (journal->j_running_transaction) {
1721                 transaction = journal->j_running_transaction;
1722                 __jbd2_log_start_commit(journal, transaction->t_tid);
1723         } else if (journal->j_committing_transaction)
1724                 transaction = journal->j_committing_transaction;
1725
1726         /* Wait for the log commit to complete... */
1727         if (transaction) {
1728                 tid_t tid = transaction->t_tid;
1729
1730                 write_unlock(&journal->j_state_lock);
1731                 jbd2_log_wait_commit(journal, tid);
1732         } else {
1733                 write_unlock(&journal->j_state_lock);
1734         }
1735
1736         /* ...and flush everything in the log out to disk. */
1737         spin_lock(&journal->j_list_lock);
1738         while (!err && journal->j_checkpoint_transactions != NULL) {
1739                 spin_unlock(&journal->j_list_lock);
1740                 mutex_lock(&journal->j_checkpoint_mutex);
1741                 err = jbd2_log_do_checkpoint(journal);
1742                 mutex_unlock(&journal->j_checkpoint_mutex);
1743                 spin_lock(&journal->j_list_lock);
1744         }
1745         spin_unlock(&journal->j_list_lock);
1746
1747         if (is_journal_aborted(journal))
1748                 return -EIO;
1749
1750         mutex_lock(&journal->j_checkpoint_mutex);
1751         jbd2_cleanup_journal_tail(journal);
1752
1753         /* Finally, mark the journal as really needing no recovery.
1754          * This sets s_start==0 in the underlying superblock, which is
1755          * the magic code for a fully-recovered superblock.  Any future
1756          * commits of data to the journal will restore the current
1757          * s_start value. */
1758         jbd2_mark_journal_empty(journal);
1759         mutex_unlock(&journal->j_checkpoint_mutex);
1760         write_lock(&journal->j_state_lock);
1761         J_ASSERT(!journal->j_running_transaction);
1762         J_ASSERT(!journal->j_committing_transaction);
1763         J_ASSERT(!journal->j_checkpoint_transactions);
1764         J_ASSERT(journal->j_head == journal->j_tail);
1765         J_ASSERT(journal->j_tail_sequence == journal->j_transaction_sequence);
1766         write_unlock(&journal->j_state_lock);
1767         return 0;
1768 }
1769
1770 /**
1771  * int jbd2_journal_wipe() - Wipe journal contents
1772  * @journal: Journal to act on.
1773  * @write: flag (see below)
1774  *
1775  * Wipe out all of the contents of a journal, safely.  This will produce
1776  * a warning if the journal contains any valid recovery information.
1777  * Must be called between journal_init_*() and jbd2_journal_load().
1778  *
1779  * If 'write' is non-zero, then we wipe out the journal on disk; otherwise
1780  * we merely suppress recovery.
1781  */
1782
1783 int jbd2_journal_wipe(journal_t *journal, int write)
1784 {
1785         int err = 0;
1786
1787         J_ASSERT (!(journal->j_flags & JBD2_LOADED));
1788
1789         err = load_superblock(journal);
1790         if (err)
1791                 return err;
1792
1793         if (!journal->j_tail)
1794                 goto no_recovery;
1795
1796         printk(KERN_WARNING "JBD2: %s recovery information on journal\n",
1797                 write ? "Clearing" : "Ignoring");
1798
1799         err = jbd2_journal_skip_recovery(journal);
1800         if (write) {
1801                 /* Lock to make assertions happy... */
1802                 mutex_lock(&journal->j_checkpoint_mutex);
1803                 jbd2_mark_journal_empty(journal);
1804                 mutex_unlock(&journal->j_checkpoint_mutex);
1805         }
1806
1807  no_recovery:
1808         return err;
1809 }
1810
1811 /*
1812  * Journal abort has very specific semantics, which we describe
1813  * for journal abort.
1814  *
1815  * Two internal functions, which provide abort to the jbd layer
1816  * itself are here.
1817  */
1818
1819 /*
1820  * Quick version for internal journal use (doesn't lock the journal).
1821  * Aborts hard --- we mark the abort as occurred, but do _nothing_ else,
1822  * and don't attempt to make any other journal updates.
1823  */
1824 void __jbd2_journal_abort_hard(journal_t *journal)
1825 {
1826         transaction_t *transaction;
1827
1828         if (journal->j_flags & JBD2_ABORT)
1829                 return;
1830
1831         printk(KERN_ERR "Aborting journal on device %s.\n",
1832                journal->j_devname);
1833
1834         write_lock(&journal->j_state_lock);
1835         journal->j_flags |= JBD2_ABORT;
1836         transaction = journal->j_running_transaction;
1837         if (transaction)
1838                 __jbd2_log_start_commit(journal, transaction->t_tid);
1839         write_unlock(&journal->j_state_lock);
1840 }
1841
1842 /* Soft abort: record the abort error status in the journal superblock,
1843  * but don't do any other IO. */
1844 static void __journal_abort_soft (journal_t *journal, int errno)
1845 {
1846         if (journal->j_flags & JBD2_ABORT)
1847                 return;
1848
1849         if (!journal->j_errno)
1850                 journal->j_errno = errno;
1851
1852         __jbd2_journal_abort_hard(journal);
1853
1854         if (errno)
1855                 jbd2_journal_update_sb_errno(journal);
1856 }
1857
1858 /**
1859  * void jbd2_journal_abort () - Shutdown the journal immediately.
1860  * @journal: the journal to shutdown.
1861  * @errno:   an error number to record in the journal indicating
1862  *           the reason for the shutdown.
1863  *
1864  * Perform a complete, immediate shutdown of the ENTIRE
1865  * journal (not of a single transaction).  This operation cannot be
1866  * undone without closing and reopening the journal.
1867  *
1868  * The jbd2_journal_abort function is intended to support higher level error
1869  * recovery mechanisms such as the ext2/ext3 remount-readonly error
1870  * mode.
1871  *
1872  * Journal abort has very specific semantics.  Any existing dirty,
1873  * unjournaled buffers in the main filesystem will still be written to
1874  * disk by bdflush, but the journaling mechanism will be suspended
1875  * immediately and no further transaction commits will be honoured.
1876  *
1877  * Any dirty, journaled buffers will be written back to disk without
1878  * hitting the journal.  Atomicity cannot be guaranteed on an aborted
1879  * filesystem, but we _do_ attempt to leave as much data as possible
1880  * behind for fsck to use for cleanup.
1881  *
1882  * Any attempt to get a new transaction handle on a journal which is in
1883  * ABORT state will just result in an -EROFS error return.  A
1884  * jbd2_journal_stop on an existing handle will return -EIO if we have
1885  * entered abort state during the update.
1886  *
1887  * Recursive transactions are not disturbed by journal abort until the
1888  * final jbd2_journal_stop, which will receive the -EIO error.
1889  *
1890  * Finally, the jbd2_journal_abort call allows the caller to supply an errno
1891  * which will be recorded (if possible) in the journal superblock.  This
1892  * allows a client to record failure conditions in the middle of a
1893  * transaction without having to complete the transaction to record the
1894  * failure to disk.  ext3_error, for example, now uses this
1895  * functionality.
1896  *
1897  * Errors which originate from within the journaling layer will NOT
1898  * supply an errno; a null errno implies that absolutely no further
1899  * writes are done to the journal (unless there are any already in
1900  * progress).
1901  *
1902  */
1903
1904 void jbd2_journal_abort(journal_t *journal, int errno)
1905 {
1906         __journal_abort_soft(journal, errno);
1907 }
1908
1909 /**
1910  * int jbd2_journal_errno () - returns the journal's error state.
1911  * @journal: journal to examine.
1912  *
1913  * This is the errno number set with jbd2_journal_abort(), the last
1914  * time the journal was mounted - if the journal was stopped
1915  * without calling abort this will be 0.
1916  *
1917  * If the journal has been aborted on this mount time -EROFS will
1918  * be returned.
1919  */
1920 int jbd2_journal_errno(journal_t *journal)
1921 {
1922         int err;
1923
1924         read_lock(&journal->j_state_lock);
1925         if (journal->j_flags & JBD2_ABORT)
1926                 err = -EROFS;
1927         else
1928                 err = journal->j_errno;
1929         read_unlock(&journal->j_state_lock);
1930         return err;
1931 }
1932
1933 /**
1934  * int jbd2_journal_clear_err () - clears the journal's error state
1935  * @journal: journal to act on.
1936  *
1937  * An error must be cleared or acked to take a FS out of readonly
1938  * mode.
1939  */
1940 int jbd2_journal_clear_err(journal_t *journal)
1941 {
1942         int err = 0;
1943
1944         write_lock(&journal->j_state_lock);
1945         if (journal->j_flags & JBD2_ABORT)
1946                 err = -EROFS;
1947         else
1948                 journal->j_errno = 0;
1949         write_unlock(&journal->j_state_lock);
1950         return err;
1951 }
1952
1953 /**
1954  * void jbd2_journal_ack_err() - Ack journal err.
1955  * @journal: journal to act on.
1956  *
1957  * An error must be cleared or acked to take a FS out of readonly
1958  * mode.
1959  */
1960 void jbd2_journal_ack_err(journal_t *journal)
1961 {
1962         write_lock(&journal->j_state_lock);
1963         if (journal->j_errno)
1964                 journal->j_flags |= JBD2_ACK_ERR;
1965         write_unlock(&journal->j_state_lock);
1966 }
1967
1968 int jbd2_journal_blocks_per_page(struct inode *inode)
1969 {
1970         return 1 << (PAGE_CACHE_SHIFT - inode->i_sb->s_blocksize_bits);
1971 }
1972
1973 /*
1974  * helper functions to deal with 32 or 64bit block numbers.
1975  */
1976 size_t journal_tag_bytes(journal_t *journal)
1977 {
1978         if (JBD2_HAS_INCOMPAT_FEATURE(journal, JBD2_FEATURE_INCOMPAT_64BIT))
1979                 return JBD2_TAG_SIZE64;
1980         else
1981                 return JBD2_TAG_SIZE32;
1982 }
1983
1984 /*
1985  * JBD memory management
1986  *
1987  * These functions are used to allocate block-sized chunks of memory
1988  * used for making copies of buffer_head data.  Very often it will be
1989  * page-sized chunks of data, but sometimes it will be in
1990  * sub-page-size chunks.  (For example, 16k pages on Power systems
1991  * with a 4k block file system.)  For blocks smaller than a page, we
1992  * use a SLAB allocator.  There are slab caches for each block size,
1993  * which are allocated at mount time, if necessary, and we only free
1994  * (all of) the slab caches when/if the jbd2 module is unloaded.  For
1995  * this reason we don't need to a mutex to protect access to
1996  * jbd2_slab[] allocating or releasing memory; only in
1997  * jbd2_journal_create_slab().
1998  */
1999 #define JBD2_MAX_SLABS 8
2000 static struct kmem_cache *jbd2_slab[JBD2_MAX_SLABS];
2001
2002 static const char *jbd2_slab_names[JBD2_MAX_SLABS] = {
2003         "jbd2_1k", "jbd2_2k", "jbd2_4k", "jbd2_8k",
2004         "jbd2_16k", "jbd2_32k", "jbd2_64k", "jbd2_128k"
2005 };
2006
2007
2008 static void jbd2_journal_destroy_slabs(void)
2009 {
2010         int i;
2011
2012         for (i = 0; i < JBD2_MAX_SLABS; i++) {
2013                 if (jbd2_slab[i])
2014                         kmem_cache_destroy(jbd2_slab[i]);
2015                 jbd2_slab[i] = NULL;
2016         }
2017 }
2018
2019 static int jbd2_journal_create_slab(size_t size)
2020 {
2021         static DEFINE_MUTEX(jbd2_slab_create_mutex);
2022         int i = order_base_2(size) - 10;
2023         size_t slab_size;
2024
2025         if (size == PAGE_SIZE)
2026                 return 0;
2027
2028         if (i >= JBD2_MAX_SLABS)
2029                 return -EINVAL;
2030
2031         if (unlikely(i < 0))
2032                 i = 0;
2033         mutex_lock(&jbd2_slab_create_mutex);
2034         if (jbd2_slab[i]) {
2035                 mutex_unlock(&jbd2_slab_create_mutex);
2036                 return 0;       /* Already created */
2037         }
2038
2039         slab_size = 1 << (i+10);
2040         jbd2_slab[i] = kmem_cache_create(jbd2_slab_names[i], slab_size,
2041                                          slab_size, 0, NULL);
2042         mutex_unlock(&jbd2_slab_create_mutex);
2043         if (!jbd2_slab[i]) {
2044                 printk(KERN_EMERG "JBD2: no memory for jbd2_slab cache\n");
2045                 return -ENOMEM;
2046         }
2047         return 0;
2048 }
2049
2050 static struct kmem_cache *get_slab(size_t size)
2051 {
2052         int i = order_base_2(size) - 10;
2053
2054         BUG_ON(i >= JBD2_MAX_SLABS);
2055         if (unlikely(i < 0))
2056                 i = 0;
2057         BUG_ON(jbd2_slab[i] == NULL);
2058         return jbd2_slab[i];
2059 }
2060
2061 void *jbd2_alloc(size_t size, gfp_t flags)
2062 {
2063         void *ptr;
2064
2065         BUG_ON(size & (size-1)); /* Must be a power of 2 */
2066
2067         flags |= __GFP_REPEAT;
2068         if (size == PAGE_SIZE)
2069                 ptr = (void *)__get_free_pages(flags, 0);
2070         else if (size > PAGE_SIZE) {
2071                 int order = get_order(size);
2072
2073                 if (order < 3)
2074                         ptr = (void *)__get_free_pages(flags, order);
2075                 else
2076                         ptr = vmalloc(size);
2077         } else
2078                 ptr = kmem_cache_alloc(get_slab(size), flags);
2079
2080         /* Check alignment; SLUB has gotten this wrong in the past,
2081          * and this can lead to user data corruption! */
2082         BUG_ON(((unsigned long) ptr) & (size-1));
2083
2084         return ptr;
2085 }
2086
2087 void jbd2_free(void *ptr, size_t size)
2088 {
2089         if (size == PAGE_SIZE) {
2090                 free_pages((unsigned long)ptr, 0);
2091                 return;
2092         }
2093         if (size > PAGE_SIZE) {
2094                 int order = get_order(size);
2095
2096                 if (order < 3)
2097                         free_pages((unsigned long)ptr, order);
2098                 else
2099                         vfree(ptr);
2100                 return;
2101         }
2102         kmem_cache_free(get_slab(size), ptr);
2103 };
2104
2105 /*
2106  * Journal_head storage management
2107  */
2108 static struct kmem_cache *jbd2_journal_head_cache;
2109 #ifdef CONFIG_JBD2_DEBUG
2110 static atomic_t nr_journal_heads = ATOMIC_INIT(0);
2111 #endif
2112
2113 static int jbd2_journal_init_journal_head_cache(void)
2114 {
2115         int retval;
2116
2117         J_ASSERT(jbd2_journal_head_cache == NULL);
2118         jbd2_journal_head_cache = kmem_cache_create("jbd2_journal_head",
2119                                 sizeof(struct journal_head),
2120                                 0,              /* offset */
2121                                 SLAB_TEMPORARY, /* flags */
2122                                 NULL);          /* ctor */
2123         retval = 0;
2124         if (!jbd2_journal_head_cache) {
2125                 retval = -ENOMEM;
2126                 printk(KERN_EMERG "JBD2: no memory for journal_head cache\n");
2127         }
2128         return retval;
2129 }
2130
2131 static void jbd2_journal_destroy_journal_head_cache(void)
2132 {
2133         if (jbd2_journal_head_cache) {
2134                 kmem_cache_destroy(jbd2_journal_head_cache);
2135                 jbd2_journal_head_cache = NULL;
2136         }
2137 }
2138
2139 /*
2140  * journal_head splicing and dicing
2141  */
2142 static struct journal_head *journal_alloc_journal_head(void)
2143 {
2144         struct journal_head *ret;
2145
2146 #ifdef CONFIG_JBD2_DEBUG
2147         atomic_inc(&nr_journal_heads);
2148 #endif
2149         ret = kmem_cache_alloc(jbd2_journal_head_cache, GFP_NOFS);
2150         if (!ret) {
2151                 jbd_debug(1, "out of memory for journal_head\n");
2152                 pr_notice_ratelimited("ENOMEM in %s, retrying.\n", __func__);
2153                 while (!ret) {
2154                         yield();
2155                         ret = kmem_cache_alloc(jbd2_journal_head_cache, GFP_NOFS);
2156                 }
2157         }
2158         return ret;
2159 }
2160
2161 static void journal_free_journal_head(struct journal_head *jh)
2162 {
2163 #ifdef CONFIG_JBD2_DEBUG
2164         atomic_dec(&nr_journal_heads);
2165         memset(jh, JBD2_POISON_FREE, sizeof(*jh));
2166 #endif
2167         kmem_cache_free(jbd2_journal_head_cache, jh);
2168 }
2169
2170 /*
2171  * A journal_head is attached to a buffer_head whenever JBD has an
2172  * interest in the buffer.
2173  *
2174  * Whenever a buffer has an attached journal_head, its ->b_state:BH_JBD bit
2175  * is set.  This bit is tested in core kernel code where we need to take
2176  * JBD-specific actions.  Testing the zeroness of ->b_private is not reliable
2177  * there.
2178  *
2179  * When a buffer has its BH_JBD bit set, its ->b_count is elevated by one.
2180  *
2181  * When a buffer has its BH_JBD bit set it is immune from being released by
2182  * core kernel code, mainly via ->b_count.
2183  *
2184  * A journal_head is detached from its buffer_head when the journal_head's
2185  * b_jcount reaches zero. Running transaction (b_transaction) and checkpoint
2186  * transaction (b_cp_transaction) hold their references to b_jcount.
2187  *
2188  * Various places in the kernel want to attach a journal_head to a buffer_head
2189  * _before_ attaching the journal_head to a transaction.  To protect the
2190  * journal_head in this situation, jbd2_journal_add_journal_head elevates the
2191  * journal_head's b_jcount refcount by one.  The caller must call
2192  * jbd2_journal_put_journal_head() to undo this.
2193  *
2194  * So the typical usage would be:
2195  *
2196  *      (Attach a journal_head if needed.  Increments b_jcount)
2197  *      struct journal_head *jh = jbd2_journal_add_journal_head(bh);
2198  *      ...
2199  *      (Get another reference for transaction)
2200  *      jbd2_journal_grab_journal_head(bh);
2201  *      jh->b_transaction = xxx;
2202  *      (Put original reference)
2203  *      jbd2_journal_put_journal_head(jh);
2204  */
2205
2206 /*
2207  * Give a buffer_head a journal_head.
2208  *
2209  * May sleep.
2210  */
2211 struct journal_head *jbd2_journal_add_journal_head(struct buffer_head *bh)
2212 {
2213         struct journal_head *jh;
2214         struct journal_head *new_jh = NULL;
2215
2216 repeat:
2217         if (!buffer_jbd(bh)) {
2218                 new_jh = journal_alloc_journal_head();
2219                 memset(new_jh, 0, sizeof(*new_jh));
2220         }
2221
2222         jbd_lock_bh_journal_head(bh);
2223         if (buffer_jbd(bh)) {
2224                 jh = bh2jh(bh);
2225         } else {
2226                 J_ASSERT_BH(bh,
2227                         (atomic_read(&bh->b_count) > 0) ||
2228                         (bh->b_page && bh->b_page->mapping));
2229
2230                 if (!new_jh) {
2231                         jbd_unlock_bh_journal_head(bh);
2232                         goto repeat;
2233                 }
2234
2235                 jh = new_jh;
2236                 new_jh = NULL;          /* We consumed it */
2237                 set_buffer_jbd(bh);
2238                 bh->b_private = jh;
2239                 jh->b_bh = bh;
2240                 get_bh(bh);
2241                 BUFFER_TRACE(bh, "added journal_head");
2242         }
2243         jh->b_jcount++;
2244         jbd_unlock_bh_journal_head(bh);
2245         if (new_jh)
2246                 journal_free_journal_head(new_jh);
2247         return bh->b_private;
2248 }
2249
2250 /*
2251  * Grab a ref against this buffer_head's journal_head.  If it ended up not
2252  * having a journal_head, return NULL
2253  */
2254 struct journal_head *jbd2_journal_grab_journal_head(struct buffer_head *bh)
2255 {
2256         struct journal_head *jh = NULL;
2257
2258         jbd_lock_bh_journal_head(bh);
2259         if (buffer_jbd(bh)) {
2260                 jh = bh2jh(bh);
2261                 jh->b_jcount++;
2262         }
2263         jbd_unlock_bh_journal_head(bh);
2264         return jh;
2265 }
2266
2267 static void __journal_remove_journal_head(struct buffer_head *bh)
2268 {
2269         struct journal_head *jh = bh2jh(bh);
2270
2271         J_ASSERT_JH(jh, jh->b_jcount >= 0);
2272         J_ASSERT_JH(jh, jh->b_transaction == NULL);
2273         J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
2274         J_ASSERT_JH(jh, jh->b_cp_transaction == NULL);
2275         J_ASSERT_JH(jh, jh->b_jlist == BJ_None);
2276         J_ASSERT_BH(bh, buffer_jbd(bh));
2277         J_ASSERT_BH(bh, jh2bh(jh) == bh);
2278         BUFFER_TRACE(bh, "remove journal_head");
2279         if (jh->b_frozen_data) {
2280                 printk(KERN_WARNING "%s: freeing b_frozen_data\n", __func__);
2281                 jbd2_free(jh->b_frozen_data, bh->b_size);
2282         }
2283         if (jh->b_committed_data) {
2284                 printk(KERN_WARNING "%s: freeing b_committed_data\n", __func__);
2285                 jbd2_free(jh->b_committed_data, bh->b_size);
2286         }
2287         bh->b_private = NULL;
2288         jh->b_bh = NULL;        /* debug, really */
2289         clear_buffer_jbd(bh);
2290         journal_free_journal_head(jh);
2291 }
2292
2293 /*
2294  * Drop a reference on the passed journal_head.  If it fell to zero then
2295  * release the journal_head from the buffer_head.
2296  */
2297 void jbd2_journal_put_journal_head(struct journal_head *jh)
2298 {
2299         struct buffer_head *bh = jh2bh(jh);
2300
2301         jbd_lock_bh_journal_head(bh);
2302         J_ASSERT_JH(jh, jh->b_jcount > 0);
2303         --jh->b_jcount;
2304         if (!jh->b_jcount) {
2305                 __journal_remove_journal_head(bh);
2306                 jbd_unlock_bh_journal_head(bh);
2307                 __brelse(bh);
2308         } else
2309                 jbd_unlock_bh_journal_head(bh);
2310 }
2311
2312 /*
2313  * Initialize jbd inode head
2314  */
2315 void jbd2_journal_init_jbd_inode(struct jbd2_inode *jinode, struct inode *inode)
2316 {
2317         jinode->i_transaction = NULL;
2318         jinode->i_next_transaction = NULL;
2319         jinode->i_vfs_inode = inode;
2320         jinode->i_flags = 0;
2321         INIT_LIST_HEAD(&jinode->i_list);
2322 }
2323
2324 /*
2325  * Function to be called before we start removing inode from memory (i.e.,
2326  * clear_inode() is a fine place to be called from). It removes inode from
2327  * transaction's lists.
2328  */
2329 void jbd2_journal_release_jbd_inode(journal_t *journal,
2330                                     struct jbd2_inode *jinode)
2331 {
2332         if (!journal)
2333                 return;
2334 restart:
2335         spin_lock(&journal->j_list_lock);
2336         /* Is commit writing out inode - we have to wait */
2337         if (test_bit(__JI_COMMIT_RUNNING, &jinode->i_flags)) {
2338                 wait_queue_head_t *wq;
2339                 DEFINE_WAIT_BIT(wait, &jinode->i_flags, __JI_COMMIT_RUNNING);
2340                 wq = bit_waitqueue(&jinode->i_flags, __JI_COMMIT_RUNNING);
2341                 prepare_to_wait(wq, &wait.wait, TASK_UNINTERRUPTIBLE);
2342                 spin_unlock(&journal->j_list_lock);
2343                 schedule();
2344                 finish_wait(wq, &wait.wait);
2345                 goto restart;
2346         }
2347
2348         if (jinode->i_transaction) {
2349                 list_del(&jinode->i_list);
2350                 jinode->i_transaction = NULL;
2351         }
2352         spin_unlock(&journal->j_list_lock);
2353 }
2354
2355 /*
2356  * debugfs tunables
2357  */
2358 #ifdef CONFIG_JBD2_DEBUG
2359 u8 jbd2_journal_enable_debug __read_mostly;
2360 EXPORT_SYMBOL(jbd2_journal_enable_debug);
2361
2362 #define JBD2_DEBUG_NAME "jbd2-debug"
2363
2364 static struct dentry *jbd2_debugfs_dir;
2365 static struct dentry *jbd2_debug;
2366
2367 static void __init jbd2_create_debugfs_entry(void)
2368 {
2369         jbd2_debugfs_dir = debugfs_create_dir("jbd2", NULL);
2370         if (jbd2_debugfs_dir)
2371                 jbd2_debug = debugfs_create_u8(JBD2_DEBUG_NAME,
2372                                                S_IRUGO | S_IWUSR,
2373                                                jbd2_debugfs_dir,
2374                                                &jbd2_journal_enable_debug);
2375 }
2376
2377 static void __exit jbd2_remove_debugfs_entry(void)
2378 {
2379         debugfs_remove(jbd2_debug);
2380         debugfs_remove(jbd2_debugfs_dir);
2381 }
2382
2383 #else
2384
2385 static void __init jbd2_create_debugfs_entry(void)
2386 {
2387 }
2388
2389 static void __exit jbd2_remove_debugfs_entry(void)
2390 {
2391 }
2392
2393 #endif
2394
2395 #ifdef CONFIG_PROC_FS
2396
2397 #define JBD2_STATS_PROC_NAME "fs/jbd2"
2398
2399 static void __init jbd2_create_jbd_stats_proc_entry(void)
2400 {
2401         proc_jbd2_stats = proc_mkdir(JBD2_STATS_PROC_NAME, NULL);
2402 }
2403
2404 static void __exit jbd2_remove_jbd_stats_proc_entry(void)
2405 {
2406         if (proc_jbd2_stats)
2407                 remove_proc_entry(JBD2_STATS_PROC_NAME, NULL);
2408 }
2409
2410 #else
2411
2412 #define jbd2_create_jbd_stats_proc_entry() do {} while (0)
2413 #define jbd2_remove_jbd_stats_proc_entry() do {} while (0)
2414
2415 #endif
2416
2417 struct kmem_cache *jbd2_handle_cache, *jbd2_inode_cache;
2418
2419 static int __init jbd2_journal_init_handle_cache(void)
2420 {
2421         jbd2_handle_cache = KMEM_CACHE(jbd2_journal_handle, SLAB_TEMPORARY);
2422         if (jbd2_handle_cache == NULL) {
2423                 printk(KERN_EMERG "JBD2: failed to create handle cache\n");
2424                 return -ENOMEM;
2425         }
2426         jbd2_inode_cache = KMEM_CACHE(jbd2_inode, 0);
2427         if (jbd2_inode_cache == NULL) {
2428                 printk(KERN_EMERG "JBD2: failed to create inode cache\n");
2429                 kmem_cache_destroy(jbd2_handle_cache);
2430                 return -ENOMEM;
2431         }
2432         return 0;
2433 }
2434
2435 static void jbd2_journal_destroy_handle_cache(void)
2436 {
2437         if (jbd2_handle_cache)
2438                 kmem_cache_destroy(jbd2_handle_cache);
2439         if (jbd2_inode_cache)
2440                 kmem_cache_destroy(jbd2_inode_cache);
2441
2442 }
2443
2444 /*
2445  * Module startup and shutdown
2446  */
2447
2448 static int __init journal_init_caches(void)
2449 {
2450         int ret;
2451
2452         ret = jbd2_journal_init_revoke_caches();
2453         if (ret == 0)
2454                 ret = jbd2_journal_init_journal_head_cache();
2455         if (ret == 0)
2456                 ret = jbd2_journal_init_handle_cache();
2457         if (ret == 0)
2458                 ret = jbd2_journal_init_transaction_cache();
2459         return ret;
2460 }
2461
2462 static void jbd2_journal_destroy_caches(void)
2463 {
2464         jbd2_journal_destroy_revoke_caches();
2465         jbd2_journal_destroy_journal_head_cache();
2466         jbd2_journal_destroy_handle_cache();
2467         jbd2_journal_destroy_transaction_cache();
2468         jbd2_journal_destroy_slabs();
2469 }
2470
2471 static int __init journal_init(void)
2472 {
2473         int ret;
2474
2475         BUILD_BUG_ON(sizeof(struct journal_superblock_s) != 1024);
2476
2477         ret = journal_init_caches();
2478         if (ret == 0) {
2479                 jbd2_create_debugfs_entry();
2480                 jbd2_create_jbd_stats_proc_entry();
2481         } else {
2482                 jbd2_journal_destroy_caches();
2483         }
2484         return ret;
2485 }
2486
2487 static void __exit journal_exit(void)
2488 {
2489 #ifdef CONFIG_JBD2_DEBUG
2490         int n = atomic_read(&nr_journal_heads);
2491         if (n)
2492                 printk(KERN_EMERG "JBD2: leaked %d journal_heads!\n", n);
2493 #endif
2494         jbd2_remove_debugfs_entry();
2495         jbd2_remove_jbd_stats_proc_entry();
2496         jbd2_journal_destroy_caches();
2497 }
2498
2499 MODULE_LICENSE("GPL");
2500 module_init(journal_init);
2501 module_exit(journal_exit);
2502