]> git.karo-electronics.de Git - karo-tx-linux.git/blob - fs/jbd2/journal.c
Merge tag 'md-3.3-fixes' of git://neil.brown.name/md
[karo-tx-linux.git] / fs / jbd2 / journal.c
1 /*
2  * linux/fs/jbd2/journal.c
3  *
4  * Written by Stephen C. Tweedie <sct@redhat.com>, 1998
5  *
6  * Copyright 1998 Red Hat corp --- All Rights Reserved
7  *
8  * This file is part of the Linux kernel and is made available under
9  * the terms of the GNU General Public License, version 2, or at your
10  * option, any later version, incorporated herein by reference.
11  *
12  * Generic filesystem journal-writing code; part of the ext2fs
13  * journaling system.
14  *
15  * This file manages journals: areas of disk reserved for logging
16  * transactional updates.  This includes the kernel journaling thread
17  * which is responsible for scheduling updates to the log.
18  *
19  * We do not actually manage the physical storage of the journal in this
20  * file: that is left to a per-journal policy function, which allows us
21  * to store the journal within a filesystem-specified area for ext2
22  * journaling (ext2 can use a reserved inode for storing the log).
23  */
24
25 #include <linux/module.h>
26 #include <linux/time.h>
27 #include <linux/fs.h>
28 #include <linux/jbd2.h>
29 #include <linux/errno.h>
30 #include <linux/slab.h>
31 #include <linux/init.h>
32 #include <linux/mm.h>
33 #include <linux/freezer.h>
34 #include <linux/pagemap.h>
35 #include <linux/kthread.h>
36 #include <linux/poison.h>
37 #include <linux/proc_fs.h>
38 #include <linux/debugfs.h>
39 #include <linux/seq_file.h>
40 #include <linux/math64.h>
41 #include <linux/hash.h>
42 #include <linux/log2.h>
43 #include <linux/vmalloc.h>
44 #include <linux/backing-dev.h>
45 #include <linux/bitops.h>
46 #include <linux/ratelimit.h>
47
48 #define CREATE_TRACE_POINTS
49 #include <trace/events/jbd2.h>
50
51 #include <asm/uaccess.h>
52 #include <asm/page.h>
53 #include <asm/system.h>
54
55 EXPORT_SYMBOL(jbd2_journal_extend);
56 EXPORT_SYMBOL(jbd2_journal_stop);
57 EXPORT_SYMBOL(jbd2_journal_lock_updates);
58 EXPORT_SYMBOL(jbd2_journal_unlock_updates);
59 EXPORT_SYMBOL(jbd2_journal_get_write_access);
60 EXPORT_SYMBOL(jbd2_journal_get_create_access);
61 EXPORT_SYMBOL(jbd2_journal_get_undo_access);
62 EXPORT_SYMBOL(jbd2_journal_set_triggers);
63 EXPORT_SYMBOL(jbd2_journal_dirty_metadata);
64 EXPORT_SYMBOL(jbd2_journal_release_buffer);
65 EXPORT_SYMBOL(jbd2_journal_forget);
66 #if 0
67 EXPORT_SYMBOL(journal_sync_buffer);
68 #endif
69 EXPORT_SYMBOL(jbd2_journal_flush);
70 EXPORT_SYMBOL(jbd2_journal_revoke);
71
72 EXPORT_SYMBOL(jbd2_journal_init_dev);
73 EXPORT_SYMBOL(jbd2_journal_init_inode);
74 EXPORT_SYMBOL(jbd2_journal_update_format);
75 EXPORT_SYMBOL(jbd2_journal_check_used_features);
76 EXPORT_SYMBOL(jbd2_journal_check_available_features);
77 EXPORT_SYMBOL(jbd2_journal_set_features);
78 EXPORT_SYMBOL(jbd2_journal_load);
79 EXPORT_SYMBOL(jbd2_journal_destroy);
80 EXPORT_SYMBOL(jbd2_journal_abort);
81 EXPORT_SYMBOL(jbd2_journal_errno);
82 EXPORT_SYMBOL(jbd2_journal_ack_err);
83 EXPORT_SYMBOL(jbd2_journal_clear_err);
84 EXPORT_SYMBOL(jbd2_log_wait_commit);
85 EXPORT_SYMBOL(jbd2_log_start_commit);
86 EXPORT_SYMBOL(jbd2_journal_start_commit);
87 EXPORT_SYMBOL(jbd2_journal_force_commit_nested);
88 EXPORT_SYMBOL(jbd2_journal_wipe);
89 EXPORT_SYMBOL(jbd2_journal_blocks_per_page);
90 EXPORT_SYMBOL(jbd2_journal_invalidatepage);
91 EXPORT_SYMBOL(jbd2_journal_try_to_free_buffers);
92 EXPORT_SYMBOL(jbd2_journal_force_commit);
93 EXPORT_SYMBOL(jbd2_journal_file_inode);
94 EXPORT_SYMBOL(jbd2_journal_init_jbd_inode);
95 EXPORT_SYMBOL(jbd2_journal_release_jbd_inode);
96 EXPORT_SYMBOL(jbd2_journal_begin_ordered_truncate);
97 EXPORT_SYMBOL(jbd2_inode_cache);
98
99 static int journal_convert_superblock_v1(journal_t *, journal_superblock_t *);
100 static void __journal_abort_soft (journal_t *journal, int errno);
101 static int jbd2_journal_create_slab(size_t slab_size);
102
103 /*
104  * Helper function used to manage commit timeouts
105  */
106
107 static void commit_timeout(unsigned long __data)
108 {
109         struct task_struct * p = (struct task_struct *) __data;
110
111         wake_up_process(p);
112 }
113
114 /*
115  * kjournald2: The main thread function used to manage a logging device
116  * journal.
117  *
118  * This kernel thread is responsible for two things:
119  *
120  * 1) COMMIT:  Every so often we need to commit the current state of the
121  *    filesystem to disk.  The journal thread is responsible for writing
122  *    all of the metadata buffers to disk.
123  *
124  * 2) CHECKPOINT: We cannot reuse a used section of the log file until all
125  *    of the data in that part of the log has been rewritten elsewhere on
126  *    the disk.  Flushing these old buffers to reclaim space in the log is
127  *    known as checkpointing, and this thread is responsible for that job.
128  */
129
130 static int kjournald2(void *arg)
131 {
132         journal_t *journal = arg;
133         transaction_t *transaction;
134
135         /*
136          * Set up an interval timer which can be used to trigger a commit wakeup
137          * after the commit interval expires
138          */
139         setup_timer(&journal->j_commit_timer, commit_timeout,
140                         (unsigned long)current);
141
142         /* Record that the journal thread is running */
143         journal->j_task = current;
144         wake_up(&journal->j_wait_done_commit);
145
146         /*
147          * And now, wait forever for commit wakeup events.
148          */
149         write_lock(&journal->j_state_lock);
150
151 loop:
152         if (journal->j_flags & JBD2_UNMOUNT)
153                 goto end_loop;
154
155         jbd_debug(1, "commit_sequence=%d, commit_request=%d\n",
156                 journal->j_commit_sequence, journal->j_commit_request);
157
158         if (journal->j_commit_sequence != journal->j_commit_request) {
159                 jbd_debug(1, "OK, requests differ\n");
160                 write_unlock(&journal->j_state_lock);
161                 del_timer_sync(&journal->j_commit_timer);
162                 jbd2_journal_commit_transaction(journal);
163                 write_lock(&journal->j_state_lock);
164                 goto loop;
165         }
166
167         wake_up(&journal->j_wait_done_commit);
168         if (freezing(current)) {
169                 /*
170                  * The simpler the better. Flushing journal isn't a
171                  * good idea, because that depends on threads that may
172                  * be already stopped.
173                  */
174                 jbd_debug(1, "Now suspending kjournald2\n");
175                 write_unlock(&journal->j_state_lock);
176                 try_to_freeze();
177                 write_lock(&journal->j_state_lock);
178         } else {
179                 /*
180                  * We assume on resume that commits are already there,
181                  * so we don't sleep
182                  */
183                 DEFINE_WAIT(wait);
184                 int should_sleep = 1;
185
186                 prepare_to_wait(&journal->j_wait_commit, &wait,
187                                 TASK_INTERRUPTIBLE);
188                 if (journal->j_commit_sequence != journal->j_commit_request)
189                         should_sleep = 0;
190                 transaction = journal->j_running_transaction;
191                 if (transaction && time_after_eq(jiffies,
192                                                 transaction->t_expires))
193                         should_sleep = 0;
194                 if (journal->j_flags & JBD2_UNMOUNT)
195                         should_sleep = 0;
196                 if (should_sleep) {
197                         write_unlock(&journal->j_state_lock);
198                         schedule();
199                         write_lock(&journal->j_state_lock);
200                 }
201                 finish_wait(&journal->j_wait_commit, &wait);
202         }
203
204         jbd_debug(1, "kjournald2 wakes\n");
205
206         /*
207          * Were we woken up by a commit wakeup event?
208          */
209         transaction = journal->j_running_transaction;
210         if (transaction && time_after_eq(jiffies, transaction->t_expires)) {
211                 journal->j_commit_request = transaction->t_tid;
212                 jbd_debug(1, "woke because of timeout\n");
213         }
214         goto loop;
215
216 end_loop:
217         write_unlock(&journal->j_state_lock);
218         del_timer_sync(&journal->j_commit_timer);
219         journal->j_task = NULL;
220         wake_up(&journal->j_wait_done_commit);
221         jbd_debug(1, "Journal thread exiting.\n");
222         return 0;
223 }
224
225 static int jbd2_journal_start_thread(journal_t *journal)
226 {
227         struct task_struct *t;
228
229         t = kthread_run(kjournald2, journal, "jbd2/%s",
230                         journal->j_devname);
231         if (IS_ERR(t))
232                 return PTR_ERR(t);
233
234         wait_event(journal->j_wait_done_commit, journal->j_task != NULL);
235         return 0;
236 }
237
238 static void journal_kill_thread(journal_t *journal)
239 {
240         write_lock(&journal->j_state_lock);
241         journal->j_flags |= JBD2_UNMOUNT;
242
243         while (journal->j_task) {
244                 wake_up(&journal->j_wait_commit);
245                 write_unlock(&journal->j_state_lock);
246                 wait_event(journal->j_wait_done_commit, journal->j_task == NULL);
247                 write_lock(&journal->j_state_lock);
248         }
249         write_unlock(&journal->j_state_lock);
250 }
251
252 /*
253  * jbd2_journal_write_metadata_buffer: write a metadata buffer to the journal.
254  *
255  * Writes a metadata buffer to a given disk block.  The actual IO is not
256  * performed but a new buffer_head is constructed which labels the data
257  * to be written with the correct destination disk block.
258  *
259  * Any magic-number escaping which needs to be done will cause a
260  * copy-out here.  If the buffer happens to start with the
261  * JBD2_MAGIC_NUMBER, then we can't write it to the log directly: the
262  * magic number is only written to the log for descripter blocks.  In
263  * this case, we copy the data and replace the first word with 0, and we
264  * return a result code which indicates that this buffer needs to be
265  * marked as an escaped buffer in the corresponding log descriptor
266  * block.  The missing word can then be restored when the block is read
267  * during recovery.
268  *
269  * If the source buffer has already been modified by a new transaction
270  * since we took the last commit snapshot, we use the frozen copy of
271  * that data for IO.  If we end up using the existing buffer_head's data
272  * for the write, then we *have* to lock the buffer to prevent anyone
273  * else from using and possibly modifying it while the IO is in
274  * progress.
275  *
276  * The function returns a pointer to the buffer_heads to be used for IO.
277  *
278  * We assume that the journal has already been locked in this function.
279  *
280  * Return value:
281  *  <0: Error
282  * >=0: Finished OK
283  *
284  * On success:
285  * Bit 0 set == escape performed on the data
286  * Bit 1 set == buffer copy-out performed (kfree the data after IO)
287  */
288
289 int jbd2_journal_write_metadata_buffer(transaction_t *transaction,
290                                   struct journal_head  *jh_in,
291                                   struct journal_head **jh_out,
292                                   unsigned long long blocknr)
293 {
294         int need_copy_out = 0;
295         int done_copy_out = 0;
296         int do_escape = 0;
297         char *mapped_data;
298         struct buffer_head *new_bh;
299         struct journal_head *new_jh;
300         struct page *new_page;
301         unsigned int new_offset;
302         struct buffer_head *bh_in = jh2bh(jh_in);
303         journal_t *journal = transaction->t_journal;
304
305         /*
306          * The buffer really shouldn't be locked: only the current committing
307          * transaction is allowed to write it, so nobody else is allowed
308          * to do any IO.
309          *
310          * akpm: except if we're journalling data, and write() output is
311          * also part of a shared mapping, and another thread has
312          * decided to launch a writepage() against this buffer.
313          */
314         J_ASSERT_BH(bh_in, buffer_jbddirty(bh_in));
315
316 retry_alloc:
317         new_bh = alloc_buffer_head(GFP_NOFS);
318         if (!new_bh) {
319                 /*
320                  * Failure is not an option, but __GFP_NOFAIL is going
321                  * away; so we retry ourselves here.
322                  */
323                 congestion_wait(BLK_RW_ASYNC, HZ/50);
324                 goto retry_alloc;
325         }
326
327         /* keep subsequent assertions sane */
328         new_bh->b_state = 0;
329         init_buffer(new_bh, NULL, NULL);
330         atomic_set(&new_bh->b_count, 1);
331         new_jh = jbd2_journal_add_journal_head(new_bh); /* This sleeps */
332
333         /*
334          * If a new transaction has already done a buffer copy-out, then
335          * we use that version of the data for the commit.
336          */
337         jbd_lock_bh_state(bh_in);
338 repeat:
339         if (jh_in->b_frozen_data) {
340                 done_copy_out = 1;
341                 new_page = virt_to_page(jh_in->b_frozen_data);
342                 new_offset = offset_in_page(jh_in->b_frozen_data);
343         } else {
344                 new_page = jh2bh(jh_in)->b_page;
345                 new_offset = offset_in_page(jh2bh(jh_in)->b_data);
346         }
347
348         mapped_data = kmap_atomic(new_page, KM_USER0);
349         /*
350          * Fire data frozen trigger if data already wasn't frozen.  Do this
351          * before checking for escaping, as the trigger may modify the magic
352          * offset.  If a copy-out happens afterwards, it will have the correct
353          * data in the buffer.
354          */
355         if (!done_copy_out)
356                 jbd2_buffer_frozen_trigger(jh_in, mapped_data + new_offset,
357                                            jh_in->b_triggers);
358
359         /*
360          * Check for escaping
361          */
362         if (*((__be32 *)(mapped_data + new_offset)) ==
363                                 cpu_to_be32(JBD2_MAGIC_NUMBER)) {
364                 need_copy_out = 1;
365                 do_escape = 1;
366         }
367         kunmap_atomic(mapped_data, KM_USER0);
368
369         /*
370          * Do we need to do a data copy?
371          */
372         if (need_copy_out && !done_copy_out) {
373                 char *tmp;
374
375                 jbd_unlock_bh_state(bh_in);
376                 tmp = jbd2_alloc(bh_in->b_size, GFP_NOFS);
377                 if (!tmp) {
378                         jbd2_journal_put_journal_head(new_jh);
379                         return -ENOMEM;
380                 }
381                 jbd_lock_bh_state(bh_in);
382                 if (jh_in->b_frozen_data) {
383                         jbd2_free(tmp, bh_in->b_size);
384                         goto repeat;
385                 }
386
387                 jh_in->b_frozen_data = tmp;
388                 mapped_data = kmap_atomic(new_page, KM_USER0);
389                 memcpy(tmp, mapped_data + new_offset, jh2bh(jh_in)->b_size);
390                 kunmap_atomic(mapped_data, KM_USER0);
391
392                 new_page = virt_to_page(tmp);
393                 new_offset = offset_in_page(tmp);
394                 done_copy_out = 1;
395
396                 /*
397                  * This isn't strictly necessary, as we're using frozen
398                  * data for the escaping, but it keeps consistency with
399                  * b_frozen_data usage.
400                  */
401                 jh_in->b_frozen_triggers = jh_in->b_triggers;
402         }
403
404         /*
405          * Did we need to do an escaping?  Now we've done all the
406          * copying, we can finally do so.
407          */
408         if (do_escape) {
409                 mapped_data = kmap_atomic(new_page, KM_USER0);
410                 *((unsigned int *)(mapped_data + new_offset)) = 0;
411                 kunmap_atomic(mapped_data, KM_USER0);
412         }
413
414         set_bh_page(new_bh, new_page, new_offset);
415         new_jh->b_transaction = NULL;
416         new_bh->b_size = jh2bh(jh_in)->b_size;
417         new_bh->b_bdev = transaction->t_journal->j_dev;
418         new_bh->b_blocknr = blocknr;
419         set_buffer_mapped(new_bh);
420         set_buffer_dirty(new_bh);
421
422         *jh_out = new_jh;
423
424         /*
425          * The to-be-written buffer needs to get moved to the io queue,
426          * and the original buffer whose contents we are shadowing or
427          * copying is moved to the transaction's shadow queue.
428          */
429         JBUFFER_TRACE(jh_in, "file as BJ_Shadow");
430         spin_lock(&journal->j_list_lock);
431         __jbd2_journal_file_buffer(jh_in, transaction, BJ_Shadow);
432         spin_unlock(&journal->j_list_lock);
433         jbd_unlock_bh_state(bh_in);
434
435         JBUFFER_TRACE(new_jh, "file as BJ_IO");
436         jbd2_journal_file_buffer(new_jh, transaction, BJ_IO);
437
438         return do_escape | (done_copy_out << 1);
439 }
440
441 /*
442  * Allocation code for the journal file.  Manage the space left in the
443  * journal, so that we can begin checkpointing when appropriate.
444  */
445
446 /*
447  * __jbd2_log_space_left: Return the number of free blocks left in the journal.
448  *
449  * Called with the journal already locked.
450  *
451  * Called under j_state_lock
452  */
453
454 int __jbd2_log_space_left(journal_t *journal)
455 {
456         int left = journal->j_free;
457
458         /* assert_spin_locked(&journal->j_state_lock); */
459
460         /*
461          * Be pessimistic here about the number of those free blocks which
462          * might be required for log descriptor control blocks.
463          */
464
465 #define MIN_LOG_RESERVED_BLOCKS 32 /* Allow for rounding errors */
466
467         left -= MIN_LOG_RESERVED_BLOCKS;
468
469         if (left <= 0)
470                 return 0;
471         left -= (left >> 3);
472         return left;
473 }
474
475 /*
476  * Called with j_state_lock locked for writing.
477  * Returns true if a transaction commit was started.
478  */
479 int __jbd2_log_start_commit(journal_t *journal, tid_t target)
480 {
481         /*
482          * The only transaction we can possibly wait upon is the
483          * currently running transaction (if it exists).  Otherwise,
484          * the target tid must be an old one.
485          */
486         if (journal->j_running_transaction &&
487             journal->j_running_transaction->t_tid == target) {
488                 /*
489                  * We want a new commit: OK, mark the request and wakeup the
490                  * commit thread.  We do _not_ do the commit ourselves.
491                  */
492
493                 journal->j_commit_request = target;
494                 jbd_debug(1, "JBD2: requesting commit %d/%d\n",
495                           journal->j_commit_request,
496                           journal->j_commit_sequence);
497                 wake_up(&journal->j_wait_commit);
498                 return 1;
499         } else if (!tid_geq(journal->j_commit_request, target))
500                 /* This should never happen, but if it does, preserve
501                    the evidence before kjournald goes into a loop and
502                    increments j_commit_sequence beyond all recognition. */
503                 WARN_ONCE(1, "JBD2: bad log_start_commit: %u %u %u %u\n",
504                           journal->j_commit_request,
505                           journal->j_commit_sequence,
506                           target, journal->j_running_transaction ? 
507                           journal->j_running_transaction->t_tid : 0);
508         return 0;
509 }
510
511 int jbd2_log_start_commit(journal_t *journal, tid_t tid)
512 {
513         int ret;
514
515         write_lock(&journal->j_state_lock);
516         ret = __jbd2_log_start_commit(journal, tid);
517         write_unlock(&journal->j_state_lock);
518         return ret;
519 }
520
521 /*
522  * Force and wait upon a commit if the calling process is not within
523  * transaction.  This is used for forcing out undo-protected data which contains
524  * bitmaps, when the fs is running out of space.
525  *
526  * We can only force the running transaction if we don't have an active handle;
527  * otherwise, we will deadlock.
528  *
529  * Returns true if a transaction was started.
530  */
531 int jbd2_journal_force_commit_nested(journal_t *journal)
532 {
533         transaction_t *transaction = NULL;
534         tid_t tid;
535         int need_to_start = 0;
536
537         read_lock(&journal->j_state_lock);
538         if (journal->j_running_transaction && !current->journal_info) {
539                 transaction = journal->j_running_transaction;
540                 if (!tid_geq(journal->j_commit_request, transaction->t_tid))
541                         need_to_start = 1;
542         } else if (journal->j_committing_transaction)
543                 transaction = journal->j_committing_transaction;
544
545         if (!transaction) {
546                 read_unlock(&journal->j_state_lock);
547                 return 0;       /* Nothing to retry */
548         }
549
550         tid = transaction->t_tid;
551         read_unlock(&journal->j_state_lock);
552         if (need_to_start)
553                 jbd2_log_start_commit(journal, tid);
554         jbd2_log_wait_commit(journal, tid);
555         return 1;
556 }
557
558 /*
559  * Start a commit of the current running transaction (if any).  Returns true
560  * if a transaction is going to be committed (or is currently already
561  * committing), and fills its tid in at *ptid
562  */
563 int jbd2_journal_start_commit(journal_t *journal, tid_t *ptid)
564 {
565         int ret = 0;
566
567         write_lock(&journal->j_state_lock);
568         if (journal->j_running_transaction) {
569                 tid_t tid = journal->j_running_transaction->t_tid;
570
571                 __jbd2_log_start_commit(journal, tid);
572                 /* There's a running transaction and we've just made sure
573                  * it's commit has been scheduled. */
574                 if (ptid)
575                         *ptid = tid;
576                 ret = 1;
577         } else if (journal->j_committing_transaction) {
578                 /*
579                  * If ext3_write_super() recently started a commit, then we
580                  * have to wait for completion of that transaction
581                  */
582                 if (ptid)
583                         *ptid = journal->j_committing_transaction->t_tid;
584                 ret = 1;
585         }
586         write_unlock(&journal->j_state_lock);
587         return ret;
588 }
589
590 /*
591  * Return 1 if a given transaction has not yet sent barrier request
592  * connected with a transaction commit. If 0 is returned, transaction
593  * may or may not have sent the barrier. Used to avoid sending barrier
594  * twice in common cases.
595  */
596 int jbd2_trans_will_send_data_barrier(journal_t *journal, tid_t tid)
597 {
598         int ret = 0;
599         transaction_t *commit_trans;
600
601         if (!(journal->j_flags & JBD2_BARRIER))
602                 return 0;
603         read_lock(&journal->j_state_lock);
604         /* Transaction already committed? */
605         if (tid_geq(journal->j_commit_sequence, tid))
606                 goto out;
607         commit_trans = journal->j_committing_transaction;
608         if (!commit_trans || commit_trans->t_tid != tid) {
609                 ret = 1;
610                 goto out;
611         }
612         /*
613          * Transaction is being committed and we already proceeded to
614          * submitting a flush to fs partition?
615          */
616         if (journal->j_fs_dev != journal->j_dev) {
617                 if (!commit_trans->t_need_data_flush ||
618                     commit_trans->t_state >= T_COMMIT_DFLUSH)
619                         goto out;
620         } else {
621                 if (commit_trans->t_state >= T_COMMIT_JFLUSH)
622                         goto out;
623         }
624         ret = 1;
625 out:
626         read_unlock(&journal->j_state_lock);
627         return ret;
628 }
629 EXPORT_SYMBOL(jbd2_trans_will_send_data_barrier);
630
631 /*
632  * Wait for a specified commit to complete.
633  * The caller may not hold the journal lock.
634  */
635 int jbd2_log_wait_commit(journal_t *journal, tid_t tid)
636 {
637         int err = 0;
638
639         read_lock(&journal->j_state_lock);
640 #ifdef CONFIG_JBD2_DEBUG
641         if (!tid_geq(journal->j_commit_request, tid)) {
642                 printk(KERN_EMERG
643                        "%s: error: j_commit_request=%d, tid=%d\n",
644                        __func__, journal->j_commit_request, tid);
645         }
646 #endif
647         while (tid_gt(tid, journal->j_commit_sequence)) {
648                 jbd_debug(1, "JBD2: want %d, j_commit_sequence=%d\n",
649                                   tid, journal->j_commit_sequence);
650                 wake_up(&journal->j_wait_commit);
651                 read_unlock(&journal->j_state_lock);
652                 wait_event(journal->j_wait_done_commit,
653                                 !tid_gt(tid, journal->j_commit_sequence));
654                 read_lock(&journal->j_state_lock);
655         }
656         read_unlock(&journal->j_state_lock);
657
658         if (unlikely(is_journal_aborted(journal))) {
659                 printk(KERN_EMERG "journal commit I/O error\n");
660                 err = -EIO;
661         }
662         return err;
663 }
664
665 /*
666  * Log buffer allocation routines:
667  */
668
669 int jbd2_journal_next_log_block(journal_t *journal, unsigned long long *retp)
670 {
671         unsigned long blocknr;
672
673         write_lock(&journal->j_state_lock);
674         J_ASSERT(journal->j_free > 1);
675
676         blocknr = journal->j_head;
677         journal->j_head++;
678         journal->j_free--;
679         if (journal->j_head == journal->j_last)
680                 journal->j_head = journal->j_first;
681         write_unlock(&journal->j_state_lock);
682         return jbd2_journal_bmap(journal, blocknr, retp);
683 }
684
685 /*
686  * Conversion of logical to physical block numbers for the journal
687  *
688  * On external journals the journal blocks are identity-mapped, so
689  * this is a no-op.  If needed, we can use j_blk_offset - everything is
690  * ready.
691  */
692 int jbd2_journal_bmap(journal_t *journal, unsigned long blocknr,
693                  unsigned long long *retp)
694 {
695         int err = 0;
696         unsigned long long ret;
697
698         if (journal->j_inode) {
699                 ret = bmap(journal->j_inode, blocknr);
700                 if (ret)
701                         *retp = ret;
702                 else {
703                         printk(KERN_ALERT "%s: journal block not found "
704                                         "at offset %lu on %s\n",
705                                __func__, blocknr, journal->j_devname);
706                         err = -EIO;
707                         __journal_abort_soft(journal, err);
708                 }
709         } else {
710                 *retp = blocknr; /* +journal->j_blk_offset */
711         }
712         return err;
713 }
714
715 /*
716  * We play buffer_head aliasing tricks to write data/metadata blocks to
717  * the journal without copying their contents, but for journal
718  * descriptor blocks we do need to generate bona fide buffers.
719  *
720  * After the caller of jbd2_journal_get_descriptor_buffer() has finished modifying
721  * the buffer's contents they really should run flush_dcache_page(bh->b_page).
722  * But we don't bother doing that, so there will be coherency problems with
723  * mmaps of blockdevs which hold live JBD-controlled filesystems.
724  */
725 struct journal_head *jbd2_journal_get_descriptor_buffer(journal_t *journal)
726 {
727         struct buffer_head *bh;
728         unsigned long long blocknr;
729         int err;
730
731         err = jbd2_journal_next_log_block(journal, &blocknr);
732
733         if (err)
734                 return NULL;
735
736         bh = __getblk(journal->j_dev, blocknr, journal->j_blocksize);
737         if (!bh)
738                 return NULL;
739         lock_buffer(bh);
740         memset(bh->b_data, 0, journal->j_blocksize);
741         set_buffer_uptodate(bh);
742         unlock_buffer(bh);
743         BUFFER_TRACE(bh, "return this buffer");
744         return jbd2_journal_add_journal_head(bh);
745 }
746
747 struct jbd2_stats_proc_session {
748         journal_t *journal;
749         struct transaction_stats_s *stats;
750         int start;
751         int max;
752 };
753
754 static void *jbd2_seq_info_start(struct seq_file *seq, loff_t *pos)
755 {
756         return *pos ? NULL : SEQ_START_TOKEN;
757 }
758
759 static void *jbd2_seq_info_next(struct seq_file *seq, void *v, loff_t *pos)
760 {
761         return NULL;
762 }
763
764 static int jbd2_seq_info_show(struct seq_file *seq, void *v)
765 {
766         struct jbd2_stats_proc_session *s = seq->private;
767
768         if (v != SEQ_START_TOKEN)
769                 return 0;
770         seq_printf(seq, "%lu transaction, each up to %u blocks\n",
771                         s->stats->ts_tid,
772                         s->journal->j_max_transaction_buffers);
773         if (s->stats->ts_tid == 0)
774                 return 0;
775         seq_printf(seq, "average: \n  %ums waiting for transaction\n",
776             jiffies_to_msecs(s->stats->run.rs_wait / s->stats->ts_tid));
777         seq_printf(seq, "  %ums running transaction\n",
778             jiffies_to_msecs(s->stats->run.rs_running / s->stats->ts_tid));
779         seq_printf(seq, "  %ums transaction was being locked\n",
780             jiffies_to_msecs(s->stats->run.rs_locked / s->stats->ts_tid));
781         seq_printf(seq, "  %ums flushing data (in ordered mode)\n",
782             jiffies_to_msecs(s->stats->run.rs_flushing / s->stats->ts_tid));
783         seq_printf(seq, "  %ums logging transaction\n",
784             jiffies_to_msecs(s->stats->run.rs_logging / s->stats->ts_tid));
785         seq_printf(seq, "  %lluus average transaction commit time\n",
786                    div_u64(s->journal->j_average_commit_time, 1000));
787         seq_printf(seq, "  %lu handles per transaction\n",
788             s->stats->run.rs_handle_count / s->stats->ts_tid);
789         seq_printf(seq, "  %lu blocks per transaction\n",
790             s->stats->run.rs_blocks / s->stats->ts_tid);
791         seq_printf(seq, "  %lu logged blocks per transaction\n",
792             s->stats->run.rs_blocks_logged / s->stats->ts_tid);
793         return 0;
794 }
795
796 static void jbd2_seq_info_stop(struct seq_file *seq, void *v)
797 {
798 }
799
800 static const struct seq_operations jbd2_seq_info_ops = {
801         .start  = jbd2_seq_info_start,
802         .next   = jbd2_seq_info_next,
803         .stop   = jbd2_seq_info_stop,
804         .show   = jbd2_seq_info_show,
805 };
806
807 static int jbd2_seq_info_open(struct inode *inode, struct file *file)
808 {
809         journal_t *journal = PDE(inode)->data;
810         struct jbd2_stats_proc_session *s;
811         int rc, size;
812
813         s = kmalloc(sizeof(*s), GFP_KERNEL);
814         if (s == NULL)
815                 return -ENOMEM;
816         size = sizeof(struct transaction_stats_s);
817         s->stats = kmalloc(size, GFP_KERNEL);
818         if (s->stats == NULL) {
819                 kfree(s);
820                 return -ENOMEM;
821         }
822         spin_lock(&journal->j_history_lock);
823         memcpy(s->stats, &journal->j_stats, size);
824         s->journal = journal;
825         spin_unlock(&journal->j_history_lock);
826
827         rc = seq_open(file, &jbd2_seq_info_ops);
828         if (rc == 0) {
829                 struct seq_file *m = file->private_data;
830                 m->private = s;
831         } else {
832                 kfree(s->stats);
833                 kfree(s);
834         }
835         return rc;
836
837 }
838
839 static int jbd2_seq_info_release(struct inode *inode, struct file *file)
840 {
841         struct seq_file *seq = file->private_data;
842         struct jbd2_stats_proc_session *s = seq->private;
843         kfree(s->stats);
844         kfree(s);
845         return seq_release(inode, file);
846 }
847
848 static const struct file_operations jbd2_seq_info_fops = {
849         .owner          = THIS_MODULE,
850         .open           = jbd2_seq_info_open,
851         .read           = seq_read,
852         .llseek         = seq_lseek,
853         .release        = jbd2_seq_info_release,
854 };
855
856 static struct proc_dir_entry *proc_jbd2_stats;
857
858 static void jbd2_stats_proc_init(journal_t *journal)
859 {
860         journal->j_proc_entry = proc_mkdir(journal->j_devname, proc_jbd2_stats);
861         if (journal->j_proc_entry) {
862                 proc_create_data("info", S_IRUGO, journal->j_proc_entry,
863                                  &jbd2_seq_info_fops, journal);
864         }
865 }
866
867 static void jbd2_stats_proc_exit(journal_t *journal)
868 {
869         remove_proc_entry("info", journal->j_proc_entry);
870         remove_proc_entry(journal->j_devname, proc_jbd2_stats);
871 }
872
873 /*
874  * Management for journal control blocks: functions to create and
875  * destroy journal_t structures, and to initialise and read existing
876  * journal blocks from disk.  */
877
878 /* First: create and setup a journal_t object in memory.  We initialise
879  * very few fields yet: that has to wait until we have created the
880  * journal structures from from scratch, or loaded them from disk. */
881
882 static journal_t * journal_init_common (void)
883 {
884         journal_t *journal;
885         int err;
886
887         journal = kzalloc(sizeof(*journal), GFP_KERNEL);
888         if (!journal)
889                 return NULL;
890
891         init_waitqueue_head(&journal->j_wait_transaction_locked);
892         init_waitqueue_head(&journal->j_wait_logspace);
893         init_waitqueue_head(&journal->j_wait_done_commit);
894         init_waitqueue_head(&journal->j_wait_checkpoint);
895         init_waitqueue_head(&journal->j_wait_commit);
896         init_waitqueue_head(&journal->j_wait_updates);
897         mutex_init(&journal->j_barrier);
898         mutex_init(&journal->j_checkpoint_mutex);
899         spin_lock_init(&journal->j_revoke_lock);
900         spin_lock_init(&journal->j_list_lock);
901         rwlock_init(&journal->j_state_lock);
902
903         journal->j_commit_interval = (HZ * JBD2_DEFAULT_MAX_COMMIT_AGE);
904         journal->j_min_batch_time = 0;
905         journal->j_max_batch_time = 15000; /* 15ms */
906
907         /* The journal is marked for error until we succeed with recovery! */
908         journal->j_flags = JBD2_ABORT;
909
910         /* Set up a default-sized revoke table for the new mount. */
911         err = jbd2_journal_init_revoke(journal, JOURNAL_REVOKE_DEFAULT_HASH);
912         if (err) {
913                 kfree(journal);
914                 return NULL;
915         }
916
917         spin_lock_init(&journal->j_history_lock);
918
919         return journal;
920 }
921
922 /* jbd2_journal_init_dev and jbd2_journal_init_inode:
923  *
924  * Create a journal structure assigned some fixed set of disk blocks to
925  * the journal.  We don't actually touch those disk blocks yet, but we
926  * need to set up all of the mapping information to tell the journaling
927  * system where the journal blocks are.
928  *
929  */
930
931 /**
932  *  journal_t * jbd2_journal_init_dev() - creates and initialises a journal structure
933  *  @bdev: Block device on which to create the journal
934  *  @fs_dev: Device which hold journalled filesystem for this journal.
935  *  @start: Block nr Start of journal.
936  *  @len:  Length of the journal in blocks.
937  *  @blocksize: blocksize of journalling device
938  *
939  *  Returns: a newly created journal_t *
940  *
941  *  jbd2_journal_init_dev creates a journal which maps a fixed contiguous
942  *  range of blocks on an arbitrary block device.
943  *
944  */
945 journal_t * jbd2_journal_init_dev(struct block_device *bdev,
946                         struct block_device *fs_dev,
947                         unsigned long long start, int len, int blocksize)
948 {
949         journal_t *journal = journal_init_common();
950         struct buffer_head *bh;
951         char *p;
952         int n;
953
954         if (!journal)
955                 return NULL;
956
957         /* journal descriptor can store up to n blocks -bzzz */
958         journal->j_blocksize = blocksize;
959         journal->j_dev = bdev;
960         journal->j_fs_dev = fs_dev;
961         journal->j_blk_offset = start;
962         journal->j_maxlen = len;
963         bdevname(journal->j_dev, journal->j_devname);
964         p = journal->j_devname;
965         while ((p = strchr(p, '/')))
966                 *p = '!';
967         jbd2_stats_proc_init(journal);
968         n = journal->j_blocksize / sizeof(journal_block_tag_t);
969         journal->j_wbufsize = n;
970         journal->j_wbuf = kmalloc(n * sizeof(struct buffer_head*), GFP_KERNEL);
971         if (!journal->j_wbuf) {
972                 printk(KERN_ERR "%s: Can't allocate bhs for commit thread\n",
973                         __func__);
974                 goto out_err;
975         }
976
977         bh = __getblk(journal->j_dev, start, journal->j_blocksize);
978         if (!bh) {
979                 printk(KERN_ERR
980                        "%s: Cannot get buffer for journal superblock\n",
981                        __func__);
982                 goto out_err;
983         }
984         journal->j_sb_buffer = bh;
985         journal->j_superblock = (journal_superblock_t *)bh->b_data;
986
987         return journal;
988 out_err:
989         kfree(journal->j_wbuf);
990         jbd2_stats_proc_exit(journal);
991         kfree(journal);
992         return NULL;
993 }
994
995 /**
996  *  journal_t * jbd2_journal_init_inode () - creates a journal which maps to a inode.
997  *  @inode: An inode to create the journal in
998  *
999  * jbd2_journal_init_inode creates a journal which maps an on-disk inode as
1000  * the journal.  The inode must exist already, must support bmap() and
1001  * must have all data blocks preallocated.
1002  */
1003 journal_t * jbd2_journal_init_inode (struct inode *inode)
1004 {
1005         struct buffer_head *bh;
1006         journal_t *journal = journal_init_common();
1007         char *p;
1008         int err;
1009         int n;
1010         unsigned long long blocknr;
1011
1012         if (!journal)
1013                 return NULL;
1014
1015         journal->j_dev = journal->j_fs_dev = inode->i_sb->s_bdev;
1016         journal->j_inode = inode;
1017         bdevname(journal->j_dev, journal->j_devname);
1018         p = journal->j_devname;
1019         while ((p = strchr(p, '/')))
1020                 *p = '!';
1021         p = journal->j_devname + strlen(journal->j_devname);
1022         sprintf(p, "-%lu", journal->j_inode->i_ino);
1023         jbd_debug(1,
1024                   "journal %p: inode %s/%ld, size %Ld, bits %d, blksize %ld\n",
1025                   journal, inode->i_sb->s_id, inode->i_ino,
1026                   (long long) inode->i_size,
1027                   inode->i_sb->s_blocksize_bits, inode->i_sb->s_blocksize);
1028
1029         journal->j_maxlen = inode->i_size >> inode->i_sb->s_blocksize_bits;
1030         journal->j_blocksize = inode->i_sb->s_blocksize;
1031         jbd2_stats_proc_init(journal);
1032
1033         /* journal descriptor can store up to n blocks -bzzz */
1034         n = journal->j_blocksize / sizeof(journal_block_tag_t);
1035         journal->j_wbufsize = n;
1036         journal->j_wbuf = kmalloc(n * sizeof(struct buffer_head*), GFP_KERNEL);
1037         if (!journal->j_wbuf) {
1038                 printk(KERN_ERR "%s: Can't allocate bhs for commit thread\n",
1039                         __func__);
1040                 goto out_err;
1041         }
1042
1043         err = jbd2_journal_bmap(journal, 0, &blocknr);
1044         /* If that failed, give up */
1045         if (err) {
1046                 printk(KERN_ERR "%s: Cannot locate journal superblock\n",
1047                        __func__);
1048                 goto out_err;
1049         }
1050
1051         bh = __getblk(journal->j_dev, blocknr, journal->j_blocksize);
1052         if (!bh) {
1053                 printk(KERN_ERR
1054                        "%s: Cannot get buffer for journal superblock\n",
1055                        __func__);
1056                 goto out_err;
1057         }
1058         journal->j_sb_buffer = bh;
1059         journal->j_superblock = (journal_superblock_t *)bh->b_data;
1060
1061         return journal;
1062 out_err:
1063         kfree(journal->j_wbuf);
1064         jbd2_stats_proc_exit(journal);
1065         kfree(journal);
1066         return NULL;
1067 }
1068
1069 /*
1070  * If the journal init or create aborts, we need to mark the journal
1071  * superblock as being NULL to prevent the journal destroy from writing
1072  * back a bogus superblock.
1073  */
1074 static void journal_fail_superblock (journal_t *journal)
1075 {
1076         struct buffer_head *bh = journal->j_sb_buffer;
1077         brelse(bh);
1078         journal->j_sb_buffer = NULL;
1079 }
1080
1081 /*
1082  * Given a journal_t structure, initialise the various fields for
1083  * startup of a new journaling session.  We use this both when creating
1084  * a journal, and after recovering an old journal to reset it for
1085  * subsequent use.
1086  */
1087
1088 static int journal_reset(journal_t *journal)
1089 {
1090         journal_superblock_t *sb = journal->j_superblock;
1091         unsigned long long first, last;
1092
1093         first = be32_to_cpu(sb->s_first);
1094         last = be32_to_cpu(sb->s_maxlen);
1095         if (first + JBD2_MIN_JOURNAL_BLOCKS > last + 1) {
1096                 printk(KERN_ERR "JBD2: Journal too short (blocks %llu-%llu).\n",
1097                        first, last);
1098                 journal_fail_superblock(journal);
1099                 return -EINVAL;
1100         }
1101
1102         journal->j_first = first;
1103         journal->j_last = last;
1104
1105         journal->j_head = first;
1106         journal->j_tail = first;
1107         journal->j_free = last - first;
1108
1109         journal->j_tail_sequence = journal->j_transaction_sequence;
1110         journal->j_commit_sequence = journal->j_transaction_sequence - 1;
1111         journal->j_commit_request = journal->j_commit_sequence;
1112
1113         journal->j_max_transaction_buffers = journal->j_maxlen / 4;
1114
1115         /* Add the dynamic fields and write it to disk. */
1116         jbd2_journal_update_superblock(journal, 1);
1117         return jbd2_journal_start_thread(journal);
1118 }
1119
1120 /**
1121  * void jbd2_journal_update_superblock() - Update journal sb on disk.
1122  * @journal: The journal to update.
1123  * @wait: Set to '0' if you don't want to wait for IO completion.
1124  *
1125  * Update a journal's dynamic superblock fields and write it to disk,
1126  * optionally waiting for the IO to complete.
1127  */
1128 void jbd2_journal_update_superblock(journal_t *journal, int wait)
1129 {
1130         journal_superblock_t *sb = journal->j_superblock;
1131         struct buffer_head *bh = journal->j_sb_buffer;
1132
1133         /*
1134          * As a special case, if the on-disk copy is already marked as needing
1135          * no recovery (s_start == 0) and there are no outstanding transactions
1136          * in the filesystem, then we can safely defer the superblock update
1137          * until the next commit by setting JBD2_FLUSHED.  This avoids
1138          * attempting a write to a potential-readonly device.
1139          */
1140         if (sb->s_start == 0 && journal->j_tail_sequence ==
1141                                 journal->j_transaction_sequence) {
1142                 jbd_debug(1, "JBD2: Skipping superblock update on recovered sb "
1143                         "(start %ld, seq %d, errno %d)\n",
1144                         journal->j_tail, journal->j_tail_sequence,
1145                         journal->j_errno);
1146                 goto out;
1147         }
1148
1149         if (buffer_write_io_error(bh)) {
1150                 /*
1151                  * Oh, dear.  A previous attempt to write the journal
1152                  * superblock failed.  This could happen because the
1153                  * USB device was yanked out.  Or it could happen to
1154                  * be a transient write error and maybe the block will
1155                  * be remapped.  Nothing we can do but to retry the
1156                  * write and hope for the best.
1157                  */
1158                 printk(KERN_ERR "JBD2: previous I/O error detected "
1159                        "for journal superblock update for %s.\n",
1160                        journal->j_devname);
1161                 clear_buffer_write_io_error(bh);
1162                 set_buffer_uptodate(bh);
1163         }
1164
1165         read_lock(&journal->j_state_lock);
1166         jbd_debug(1, "JBD2: updating superblock (start %ld, seq %d, errno %d)\n",
1167                   journal->j_tail, journal->j_tail_sequence, journal->j_errno);
1168
1169         sb->s_sequence = cpu_to_be32(journal->j_tail_sequence);
1170         sb->s_start    = cpu_to_be32(journal->j_tail);
1171         sb->s_errno    = cpu_to_be32(journal->j_errno);
1172         read_unlock(&journal->j_state_lock);
1173
1174         BUFFER_TRACE(bh, "marking dirty");
1175         mark_buffer_dirty(bh);
1176         if (wait) {
1177                 sync_dirty_buffer(bh);
1178                 if (buffer_write_io_error(bh)) {
1179                         printk(KERN_ERR "JBD2: I/O error detected "
1180                                "when updating journal superblock for %s.\n",
1181                                journal->j_devname);
1182                         clear_buffer_write_io_error(bh);
1183                         set_buffer_uptodate(bh);
1184                 }
1185         } else
1186                 write_dirty_buffer(bh, WRITE);
1187
1188 out:
1189         /* If we have just flushed the log (by marking s_start==0), then
1190          * any future commit will have to be careful to update the
1191          * superblock again to re-record the true start of the log. */
1192
1193         write_lock(&journal->j_state_lock);
1194         if (sb->s_start)
1195                 journal->j_flags &= ~JBD2_FLUSHED;
1196         else
1197                 journal->j_flags |= JBD2_FLUSHED;
1198         write_unlock(&journal->j_state_lock);
1199 }
1200
1201 /*
1202  * Read the superblock for a given journal, performing initial
1203  * validation of the format.
1204  */
1205
1206 static int journal_get_superblock(journal_t *journal)
1207 {
1208         struct buffer_head *bh;
1209         journal_superblock_t *sb;
1210         int err = -EIO;
1211
1212         bh = journal->j_sb_buffer;
1213
1214         J_ASSERT(bh != NULL);
1215         if (!buffer_uptodate(bh)) {
1216                 ll_rw_block(READ, 1, &bh);
1217                 wait_on_buffer(bh);
1218                 if (!buffer_uptodate(bh)) {
1219                         printk(KERN_ERR
1220                                 "JBD2: IO error reading journal superblock\n");
1221                         goto out;
1222                 }
1223         }
1224
1225         sb = journal->j_superblock;
1226
1227         err = -EINVAL;
1228
1229         if (sb->s_header.h_magic != cpu_to_be32(JBD2_MAGIC_NUMBER) ||
1230             sb->s_blocksize != cpu_to_be32(journal->j_blocksize)) {
1231                 printk(KERN_WARNING "JBD2: no valid journal superblock found\n");
1232                 goto out;
1233         }
1234
1235         switch(be32_to_cpu(sb->s_header.h_blocktype)) {
1236         case JBD2_SUPERBLOCK_V1:
1237                 journal->j_format_version = 1;
1238                 break;
1239         case JBD2_SUPERBLOCK_V2:
1240                 journal->j_format_version = 2;
1241                 break;
1242         default:
1243                 printk(KERN_WARNING "JBD2: unrecognised superblock format ID\n");
1244                 goto out;
1245         }
1246
1247         if (be32_to_cpu(sb->s_maxlen) < journal->j_maxlen)
1248                 journal->j_maxlen = be32_to_cpu(sb->s_maxlen);
1249         else if (be32_to_cpu(sb->s_maxlen) > journal->j_maxlen) {
1250                 printk(KERN_WARNING "JBD2: journal file too short\n");
1251                 goto out;
1252         }
1253
1254         if (be32_to_cpu(sb->s_first) == 0 ||
1255             be32_to_cpu(sb->s_first) >= journal->j_maxlen) {
1256                 printk(KERN_WARNING
1257                         "JBD2: Invalid start block of journal: %u\n",
1258                         be32_to_cpu(sb->s_first));
1259                 goto out;
1260         }
1261
1262         return 0;
1263
1264 out:
1265         journal_fail_superblock(journal);
1266         return err;
1267 }
1268
1269 /*
1270  * Load the on-disk journal superblock and read the key fields into the
1271  * journal_t.
1272  */
1273
1274 static int load_superblock(journal_t *journal)
1275 {
1276         int err;
1277         journal_superblock_t *sb;
1278
1279         err = journal_get_superblock(journal);
1280         if (err)
1281                 return err;
1282
1283         sb = journal->j_superblock;
1284
1285         journal->j_tail_sequence = be32_to_cpu(sb->s_sequence);
1286         journal->j_tail = be32_to_cpu(sb->s_start);
1287         journal->j_first = be32_to_cpu(sb->s_first);
1288         journal->j_last = be32_to_cpu(sb->s_maxlen);
1289         journal->j_errno = be32_to_cpu(sb->s_errno);
1290
1291         return 0;
1292 }
1293
1294
1295 /**
1296  * int jbd2_journal_load() - Read journal from disk.
1297  * @journal: Journal to act on.
1298  *
1299  * Given a journal_t structure which tells us which disk blocks contain
1300  * a journal, read the journal from disk to initialise the in-memory
1301  * structures.
1302  */
1303 int jbd2_journal_load(journal_t *journal)
1304 {
1305         int err;
1306         journal_superblock_t *sb;
1307
1308         err = load_superblock(journal);
1309         if (err)
1310                 return err;
1311
1312         sb = journal->j_superblock;
1313         /* If this is a V2 superblock, then we have to check the
1314          * features flags on it. */
1315
1316         if (journal->j_format_version >= 2) {
1317                 if ((sb->s_feature_ro_compat &
1318                      ~cpu_to_be32(JBD2_KNOWN_ROCOMPAT_FEATURES)) ||
1319                     (sb->s_feature_incompat &
1320                      ~cpu_to_be32(JBD2_KNOWN_INCOMPAT_FEATURES))) {
1321                         printk(KERN_WARNING
1322                                 "JBD2: Unrecognised features on journal\n");
1323                         return -EINVAL;
1324                 }
1325         }
1326
1327         /*
1328          * Create a slab for this blocksize
1329          */
1330         err = jbd2_journal_create_slab(be32_to_cpu(sb->s_blocksize));
1331         if (err)
1332                 return err;
1333
1334         /* Let the recovery code check whether it needs to recover any
1335          * data from the journal. */
1336         if (jbd2_journal_recover(journal))
1337                 goto recovery_error;
1338
1339         if (journal->j_failed_commit) {
1340                 printk(KERN_ERR "JBD2: journal transaction %u on %s "
1341                        "is corrupt.\n", journal->j_failed_commit,
1342                        journal->j_devname);
1343                 return -EIO;
1344         }
1345
1346         /* OK, we've finished with the dynamic journal bits:
1347          * reinitialise the dynamic contents of the superblock in memory
1348          * and reset them on disk. */
1349         if (journal_reset(journal))
1350                 goto recovery_error;
1351
1352         journal->j_flags &= ~JBD2_ABORT;
1353         journal->j_flags |= JBD2_LOADED;
1354         return 0;
1355
1356 recovery_error:
1357         printk(KERN_WARNING "JBD2: recovery failed\n");
1358         return -EIO;
1359 }
1360
1361 /**
1362  * void jbd2_journal_destroy() - Release a journal_t structure.
1363  * @journal: Journal to act on.
1364  *
1365  * Release a journal_t structure once it is no longer in use by the
1366  * journaled object.
1367  * Return <0 if we couldn't clean up the journal.
1368  */
1369 int jbd2_journal_destroy(journal_t *journal)
1370 {
1371         int err = 0;
1372
1373         /* Wait for the commit thread to wake up and die. */
1374         journal_kill_thread(journal);
1375
1376         /* Force a final log commit */
1377         if (journal->j_running_transaction)
1378                 jbd2_journal_commit_transaction(journal);
1379
1380         /* Force any old transactions to disk */
1381
1382         /* Totally anal locking here... */
1383         spin_lock(&journal->j_list_lock);
1384         while (journal->j_checkpoint_transactions != NULL) {
1385                 spin_unlock(&journal->j_list_lock);
1386                 mutex_lock(&journal->j_checkpoint_mutex);
1387                 jbd2_log_do_checkpoint(journal);
1388                 mutex_unlock(&journal->j_checkpoint_mutex);
1389                 spin_lock(&journal->j_list_lock);
1390         }
1391
1392         J_ASSERT(journal->j_running_transaction == NULL);
1393         J_ASSERT(journal->j_committing_transaction == NULL);
1394         J_ASSERT(journal->j_checkpoint_transactions == NULL);
1395         spin_unlock(&journal->j_list_lock);
1396
1397         if (journal->j_sb_buffer) {
1398                 if (!is_journal_aborted(journal)) {
1399                         /* We can now mark the journal as empty. */
1400                         journal->j_tail = 0;
1401                         journal->j_tail_sequence =
1402                                 ++journal->j_transaction_sequence;
1403                         jbd2_journal_update_superblock(journal, 1);
1404                 } else {
1405                         err = -EIO;
1406                 }
1407                 brelse(journal->j_sb_buffer);
1408         }
1409
1410         if (journal->j_proc_entry)
1411                 jbd2_stats_proc_exit(journal);
1412         if (journal->j_inode)
1413                 iput(journal->j_inode);
1414         if (journal->j_revoke)
1415                 jbd2_journal_destroy_revoke(journal);
1416         kfree(journal->j_wbuf);
1417         kfree(journal);
1418
1419         return err;
1420 }
1421
1422
1423 /**
1424  *int jbd2_journal_check_used_features () - Check if features specified are used.
1425  * @journal: Journal to check.
1426  * @compat: bitmask of compatible features
1427  * @ro: bitmask of features that force read-only mount
1428  * @incompat: bitmask of incompatible features
1429  *
1430  * Check whether the journal uses all of a given set of
1431  * features.  Return true (non-zero) if it does.
1432  **/
1433
1434 int jbd2_journal_check_used_features (journal_t *journal, unsigned long compat,
1435                                  unsigned long ro, unsigned long incompat)
1436 {
1437         journal_superblock_t *sb;
1438
1439         if (!compat && !ro && !incompat)
1440                 return 1;
1441         /* Load journal superblock if it is not loaded yet. */
1442         if (journal->j_format_version == 0 &&
1443             journal_get_superblock(journal) != 0)
1444                 return 0;
1445         if (journal->j_format_version == 1)
1446                 return 0;
1447
1448         sb = journal->j_superblock;
1449
1450         if (((be32_to_cpu(sb->s_feature_compat) & compat) == compat) &&
1451             ((be32_to_cpu(sb->s_feature_ro_compat) & ro) == ro) &&
1452             ((be32_to_cpu(sb->s_feature_incompat) & incompat) == incompat))
1453                 return 1;
1454
1455         return 0;
1456 }
1457
1458 /**
1459  * int jbd2_journal_check_available_features() - Check feature set in journalling layer
1460  * @journal: Journal to check.
1461  * @compat: bitmask of compatible features
1462  * @ro: bitmask of features that force read-only mount
1463  * @incompat: bitmask of incompatible features
1464  *
1465  * Check whether the journaling code supports the use of
1466  * all of a given set of features on this journal.  Return true
1467  * (non-zero) if it can. */
1468
1469 int jbd2_journal_check_available_features (journal_t *journal, unsigned long compat,
1470                                       unsigned long ro, unsigned long incompat)
1471 {
1472         if (!compat && !ro && !incompat)
1473                 return 1;
1474
1475         /* We can support any known requested features iff the
1476          * superblock is in version 2.  Otherwise we fail to support any
1477          * extended sb features. */
1478
1479         if (journal->j_format_version != 2)
1480                 return 0;
1481
1482         if ((compat   & JBD2_KNOWN_COMPAT_FEATURES) == compat &&
1483             (ro       & JBD2_KNOWN_ROCOMPAT_FEATURES) == ro &&
1484             (incompat & JBD2_KNOWN_INCOMPAT_FEATURES) == incompat)
1485                 return 1;
1486
1487         return 0;
1488 }
1489
1490 /**
1491  * int jbd2_journal_set_features () - Mark a given journal feature in the superblock
1492  * @journal: Journal to act on.
1493  * @compat: bitmask of compatible features
1494  * @ro: bitmask of features that force read-only mount
1495  * @incompat: bitmask of incompatible features
1496  *
1497  * Mark a given journal feature as present on the
1498  * superblock.  Returns true if the requested features could be set.
1499  *
1500  */
1501
1502 int jbd2_journal_set_features (journal_t *journal, unsigned long compat,
1503                           unsigned long ro, unsigned long incompat)
1504 {
1505         journal_superblock_t *sb;
1506
1507         if (jbd2_journal_check_used_features(journal, compat, ro, incompat))
1508                 return 1;
1509
1510         if (!jbd2_journal_check_available_features(journal, compat, ro, incompat))
1511                 return 0;
1512
1513         jbd_debug(1, "Setting new features 0x%lx/0x%lx/0x%lx\n",
1514                   compat, ro, incompat);
1515
1516         sb = journal->j_superblock;
1517
1518         sb->s_feature_compat    |= cpu_to_be32(compat);
1519         sb->s_feature_ro_compat |= cpu_to_be32(ro);
1520         sb->s_feature_incompat  |= cpu_to_be32(incompat);
1521
1522         return 1;
1523 }
1524
1525 /*
1526  * jbd2_journal_clear_features () - Clear a given journal feature in the
1527  *                                  superblock
1528  * @journal: Journal to act on.
1529  * @compat: bitmask of compatible features
1530  * @ro: bitmask of features that force read-only mount
1531  * @incompat: bitmask of incompatible features
1532  *
1533  * Clear a given journal feature as present on the
1534  * superblock.
1535  */
1536 void jbd2_journal_clear_features(journal_t *journal, unsigned long compat,
1537                                 unsigned long ro, unsigned long incompat)
1538 {
1539         journal_superblock_t *sb;
1540
1541         jbd_debug(1, "Clear features 0x%lx/0x%lx/0x%lx\n",
1542                   compat, ro, incompat);
1543
1544         sb = journal->j_superblock;
1545
1546         sb->s_feature_compat    &= ~cpu_to_be32(compat);
1547         sb->s_feature_ro_compat &= ~cpu_to_be32(ro);
1548         sb->s_feature_incompat  &= ~cpu_to_be32(incompat);
1549 }
1550 EXPORT_SYMBOL(jbd2_journal_clear_features);
1551
1552 /**
1553  * int jbd2_journal_update_format () - Update on-disk journal structure.
1554  * @journal: Journal to act on.
1555  *
1556  * Given an initialised but unloaded journal struct, poke about in the
1557  * on-disk structure to update it to the most recent supported version.
1558  */
1559 int jbd2_journal_update_format (journal_t *journal)
1560 {
1561         journal_superblock_t *sb;
1562         int err;
1563
1564         err = journal_get_superblock(journal);
1565         if (err)
1566                 return err;
1567
1568         sb = journal->j_superblock;
1569
1570         switch (be32_to_cpu(sb->s_header.h_blocktype)) {
1571         case JBD2_SUPERBLOCK_V2:
1572                 return 0;
1573         case JBD2_SUPERBLOCK_V1:
1574                 return journal_convert_superblock_v1(journal, sb);
1575         default:
1576                 break;
1577         }
1578         return -EINVAL;
1579 }
1580
1581 static int journal_convert_superblock_v1(journal_t *journal,
1582                                          journal_superblock_t *sb)
1583 {
1584         int offset, blocksize;
1585         struct buffer_head *bh;
1586
1587         printk(KERN_WARNING
1588                 "JBD2: Converting superblock from version 1 to 2.\n");
1589
1590         /* Pre-initialise new fields to zero */
1591         offset = ((char *) &(sb->s_feature_compat)) - ((char *) sb);
1592         blocksize = be32_to_cpu(sb->s_blocksize);
1593         memset(&sb->s_feature_compat, 0, blocksize-offset);
1594
1595         sb->s_nr_users = cpu_to_be32(1);
1596         sb->s_header.h_blocktype = cpu_to_be32(JBD2_SUPERBLOCK_V2);
1597         journal->j_format_version = 2;
1598
1599         bh = journal->j_sb_buffer;
1600         BUFFER_TRACE(bh, "marking dirty");
1601         mark_buffer_dirty(bh);
1602         sync_dirty_buffer(bh);
1603         return 0;
1604 }
1605
1606
1607 /**
1608  * int jbd2_journal_flush () - Flush journal
1609  * @journal: Journal to act on.
1610  *
1611  * Flush all data for a given journal to disk and empty the journal.
1612  * Filesystems can use this when remounting readonly to ensure that
1613  * recovery does not need to happen on remount.
1614  */
1615
1616 int jbd2_journal_flush(journal_t *journal)
1617 {
1618         int err = 0;
1619         transaction_t *transaction = NULL;
1620         unsigned long old_tail;
1621
1622         write_lock(&journal->j_state_lock);
1623
1624         /* Force everything buffered to the log... */
1625         if (journal->j_running_transaction) {
1626                 transaction = journal->j_running_transaction;
1627                 __jbd2_log_start_commit(journal, transaction->t_tid);
1628         } else if (journal->j_committing_transaction)
1629                 transaction = journal->j_committing_transaction;
1630
1631         /* Wait for the log commit to complete... */
1632         if (transaction) {
1633                 tid_t tid = transaction->t_tid;
1634
1635                 write_unlock(&journal->j_state_lock);
1636                 jbd2_log_wait_commit(journal, tid);
1637         } else {
1638                 write_unlock(&journal->j_state_lock);
1639         }
1640
1641         /* ...and flush everything in the log out to disk. */
1642         spin_lock(&journal->j_list_lock);
1643         while (!err && journal->j_checkpoint_transactions != NULL) {
1644                 spin_unlock(&journal->j_list_lock);
1645                 mutex_lock(&journal->j_checkpoint_mutex);
1646                 err = jbd2_log_do_checkpoint(journal);
1647                 mutex_unlock(&journal->j_checkpoint_mutex);
1648                 spin_lock(&journal->j_list_lock);
1649         }
1650         spin_unlock(&journal->j_list_lock);
1651
1652         if (is_journal_aborted(journal))
1653                 return -EIO;
1654
1655         jbd2_cleanup_journal_tail(journal);
1656
1657         /* Finally, mark the journal as really needing no recovery.
1658          * This sets s_start==0 in the underlying superblock, which is
1659          * the magic code for a fully-recovered superblock.  Any future
1660          * commits of data to the journal will restore the current
1661          * s_start value. */
1662         write_lock(&journal->j_state_lock);
1663         old_tail = journal->j_tail;
1664         journal->j_tail = 0;
1665         write_unlock(&journal->j_state_lock);
1666         jbd2_journal_update_superblock(journal, 1);
1667         write_lock(&journal->j_state_lock);
1668         journal->j_tail = old_tail;
1669
1670         J_ASSERT(!journal->j_running_transaction);
1671         J_ASSERT(!journal->j_committing_transaction);
1672         J_ASSERT(!journal->j_checkpoint_transactions);
1673         J_ASSERT(journal->j_head == journal->j_tail);
1674         J_ASSERT(journal->j_tail_sequence == journal->j_transaction_sequence);
1675         write_unlock(&journal->j_state_lock);
1676         return 0;
1677 }
1678
1679 /**
1680  * int jbd2_journal_wipe() - Wipe journal contents
1681  * @journal: Journal to act on.
1682  * @write: flag (see below)
1683  *
1684  * Wipe out all of the contents of a journal, safely.  This will produce
1685  * a warning if the journal contains any valid recovery information.
1686  * Must be called between journal_init_*() and jbd2_journal_load().
1687  *
1688  * If 'write' is non-zero, then we wipe out the journal on disk; otherwise
1689  * we merely suppress recovery.
1690  */
1691
1692 int jbd2_journal_wipe(journal_t *journal, int write)
1693 {
1694         int err = 0;
1695
1696         J_ASSERT (!(journal->j_flags & JBD2_LOADED));
1697
1698         err = load_superblock(journal);
1699         if (err)
1700                 return err;
1701
1702         if (!journal->j_tail)
1703                 goto no_recovery;
1704
1705         printk(KERN_WARNING "JBD2: %s recovery information on journal\n",
1706                 write ? "Clearing" : "Ignoring");
1707
1708         err = jbd2_journal_skip_recovery(journal);
1709         if (write)
1710                 jbd2_journal_update_superblock(journal, 1);
1711
1712  no_recovery:
1713         return err;
1714 }
1715
1716 /*
1717  * Journal abort has very specific semantics, which we describe
1718  * for journal abort.
1719  *
1720  * Two internal functions, which provide abort to the jbd layer
1721  * itself are here.
1722  */
1723
1724 /*
1725  * Quick version for internal journal use (doesn't lock the journal).
1726  * Aborts hard --- we mark the abort as occurred, but do _nothing_ else,
1727  * and don't attempt to make any other journal updates.
1728  */
1729 void __jbd2_journal_abort_hard(journal_t *journal)
1730 {
1731         transaction_t *transaction;
1732
1733         if (journal->j_flags & JBD2_ABORT)
1734                 return;
1735
1736         printk(KERN_ERR "Aborting journal on device %s.\n",
1737                journal->j_devname);
1738
1739         write_lock(&journal->j_state_lock);
1740         journal->j_flags |= JBD2_ABORT;
1741         transaction = journal->j_running_transaction;
1742         if (transaction)
1743                 __jbd2_log_start_commit(journal, transaction->t_tid);
1744         write_unlock(&journal->j_state_lock);
1745 }
1746
1747 /* Soft abort: record the abort error status in the journal superblock,
1748  * but don't do any other IO. */
1749 static void __journal_abort_soft (journal_t *journal, int errno)
1750 {
1751         if (journal->j_flags & JBD2_ABORT)
1752                 return;
1753
1754         if (!journal->j_errno)
1755                 journal->j_errno = errno;
1756
1757         __jbd2_journal_abort_hard(journal);
1758
1759         if (errno)
1760                 jbd2_journal_update_superblock(journal, 1);
1761 }
1762
1763 /**
1764  * void jbd2_journal_abort () - Shutdown the journal immediately.
1765  * @journal: the journal to shutdown.
1766  * @errno:   an error number to record in the journal indicating
1767  *           the reason for the shutdown.
1768  *
1769  * Perform a complete, immediate shutdown of the ENTIRE
1770  * journal (not of a single transaction).  This operation cannot be
1771  * undone without closing and reopening the journal.
1772  *
1773  * The jbd2_journal_abort function is intended to support higher level error
1774  * recovery mechanisms such as the ext2/ext3 remount-readonly error
1775  * mode.
1776  *
1777  * Journal abort has very specific semantics.  Any existing dirty,
1778  * unjournaled buffers in the main filesystem will still be written to
1779  * disk by bdflush, but the journaling mechanism will be suspended
1780  * immediately and no further transaction commits will be honoured.
1781  *
1782  * Any dirty, journaled buffers will be written back to disk without
1783  * hitting the journal.  Atomicity cannot be guaranteed on an aborted
1784  * filesystem, but we _do_ attempt to leave as much data as possible
1785  * behind for fsck to use for cleanup.
1786  *
1787  * Any attempt to get a new transaction handle on a journal which is in
1788  * ABORT state will just result in an -EROFS error return.  A
1789  * jbd2_journal_stop on an existing handle will return -EIO if we have
1790  * entered abort state during the update.
1791  *
1792  * Recursive transactions are not disturbed by journal abort until the
1793  * final jbd2_journal_stop, which will receive the -EIO error.
1794  *
1795  * Finally, the jbd2_journal_abort call allows the caller to supply an errno
1796  * which will be recorded (if possible) in the journal superblock.  This
1797  * allows a client to record failure conditions in the middle of a
1798  * transaction without having to complete the transaction to record the
1799  * failure to disk.  ext3_error, for example, now uses this
1800  * functionality.
1801  *
1802  * Errors which originate from within the journaling layer will NOT
1803  * supply an errno; a null errno implies that absolutely no further
1804  * writes are done to the journal (unless there are any already in
1805  * progress).
1806  *
1807  */
1808
1809 void jbd2_journal_abort(journal_t *journal, int errno)
1810 {
1811         __journal_abort_soft(journal, errno);
1812 }
1813
1814 /**
1815  * int jbd2_journal_errno () - returns the journal's error state.
1816  * @journal: journal to examine.
1817  *
1818  * This is the errno number set with jbd2_journal_abort(), the last
1819  * time the journal was mounted - if the journal was stopped
1820  * without calling abort this will be 0.
1821  *
1822  * If the journal has been aborted on this mount time -EROFS will
1823  * be returned.
1824  */
1825 int jbd2_journal_errno(journal_t *journal)
1826 {
1827         int err;
1828
1829         read_lock(&journal->j_state_lock);
1830         if (journal->j_flags & JBD2_ABORT)
1831                 err = -EROFS;
1832         else
1833                 err = journal->j_errno;
1834         read_unlock(&journal->j_state_lock);
1835         return err;
1836 }
1837
1838 /**
1839  * int jbd2_journal_clear_err () - clears the journal's error state
1840  * @journal: journal to act on.
1841  *
1842  * An error must be cleared or acked to take a FS out of readonly
1843  * mode.
1844  */
1845 int jbd2_journal_clear_err(journal_t *journal)
1846 {
1847         int err = 0;
1848
1849         write_lock(&journal->j_state_lock);
1850         if (journal->j_flags & JBD2_ABORT)
1851                 err = -EROFS;
1852         else
1853                 journal->j_errno = 0;
1854         write_unlock(&journal->j_state_lock);
1855         return err;
1856 }
1857
1858 /**
1859  * void jbd2_journal_ack_err() - Ack journal err.
1860  * @journal: journal to act on.
1861  *
1862  * An error must be cleared or acked to take a FS out of readonly
1863  * mode.
1864  */
1865 void jbd2_journal_ack_err(journal_t *journal)
1866 {
1867         write_lock(&journal->j_state_lock);
1868         if (journal->j_errno)
1869                 journal->j_flags |= JBD2_ACK_ERR;
1870         write_unlock(&journal->j_state_lock);
1871 }
1872
1873 int jbd2_journal_blocks_per_page(struct inode *inode)
1874 {
1875         return 1 << (PAGE_CACHE_SHIFT - inode->i_sb->s_blocksize_bits);
1876 }
1877
1878 /*
1879  * helper functions to deal with 32 or 64bit block numbers.
1880  */
1881 size_t journal_tag_bytes(journal_t *journal)
1882 {
1883         if (JBD2_HAS_INCOMPAT_FEATURE(journal, JBD2_FEATURE_INCOMPAT_64BIT))
1884                 return JBD2_TAG_SIZE64;
1885         else
1886                 return JBD2_TAG_SIZE32;
1887 }
1888
1889 /*
1890  * JBD memory management
1891  *
1892  * These functions are used to allocate block-sized chunks of memory
1893  * used for making copies of buffer_head data.  Very often it will be
1894  * page-sized chunks of data, but sometimes it will be in
1895  * sub-page-size chunks.  (For example, 16k pages on Power systems
1896  * with a 4k block file system.)  For blocks smaller than a page, we
1897  * use a SLAB allocator.  There are slab caches for each block size,
1898  * which are allocated at mount time, if necessary, and we only free
1899  * (all of) the slab caches when/if the jbd2 module is unloaded.  For
1900  * this reason we don't need to a mutex to protect access to
1901  * jbd2_slab[] allocating or releasing memory; only in
1902  * jbd2_journal_create_slab().
1903  */
1904 #define JBD2_MAX_SLABS 8
1905 static struct kmem_cache *jbd2_slab[JBD2_MAX_SLABS];
1906
1907 static const char *jbd2_slab_names[JBD2_MAX_SLABS] = {
1908         "jbd2_1k", "jbd2_2k", "jbd2_4k", "jbd2_8k",
1909         "jbd2_16k", "jbd2_32k", "jbd2_64k", "jbd2_128k"
1910 };
1911
1912
1913 static void jbd2_journal_destroy_slabs(void)
1914 {
1915         int i;
1916
1917         for (i = 0; i < JBD2_MAX_SLABS; i++) {
1918                 if (jbd2_slab[i])
1919                         kmem_cache_destroy(jbd2_slab[i]);
1920                 jbd2_slab[i] = NULL;
1921         }
1922 }
1923
1924 static int jbd2_journal_create_slab(size_t size)
1925 {
1926         static DEFINE_MUTEX(jbd2_slab_create_mutex);
1927         int i = order_base_2(size) - 10;
1928         size_t slab_size;
1929
1930         if (size == PAGE_SIZE)
1931                 return 0;
1932
1933         if (i >= JBD2_MAX_SLABS)
1934                 return -EINVAL;
1935
1936         if (unlikely(i < 0))
1937                 i = 0;
1938         mutex_lock(&jbd2_slab_create_mutex);
1939         if (jbd2_slab[i]) {
1940                 mutex_unlock(&jbd2_slab_create_mutex);
1941                 return 0;       /* Already created */
1942         }
1943
1944         slab_size = 1 << (i+10);
1945         jbd2_slab[i] = kmem_cache_create(jbd2_slab_names[i], slab_size,
1946                                          slab_size, 0, NULL);
1947         mutex_unlock(&jbd2_slab_create_mutex);
1948         if (!jbd2_slab[i]) {
1949                 printk(KERN_EMERG "JBD2: no memory for jbd2_slab cache\n");
1950                 return -ENOMEM;
1951         }
1952         return 0;
1953 }
1954
1955 static struct kmem_cache *get_slab(size_t size)
1956 {
1957         int i = order_base_2(size) - 10;
1958
1959         BUG_ON(i >= JBD2_MAX_SLABS);
1960         if (unlikely(i < 0))
1961                 i = 0;
1962         BUG_ON(jbd2_slab[i] == NULL);
1963         return jbd2_slab[i];
1964 }
1965
1966 void *jbd2_alloc(size_t size, gfp_t flags)
1967 {
1968         void *ptr;
1969
1970         BUG_ON(size & (size-1)); /* Must be a power of 2 */
1971
1972         flags |= __GFP_REPEAT;
1973         if (size == PAGE_SIZE)
1974                 ptr = (void *)__get_free_pages(flags, 0);
1975         else if (size > PAGE_SIZE) {
1976                 int order = get_order(size);
1977
1978                 if (order < 3)
1979                         ptr = (void *)__get_free_pages(flags, order);
1980                 else
1981                         ptr = vmalloc(size);
1982         } else
1983                 ptr = kmem_cache_alloc(get_slab(size), flags);
1984
1985         /* Check alignment; SLUB has gotten this wrong in the past,
1986          * and this can lead to user data corruption! */
1987         BUG_ON(((unsigned long) ptr) & (size-1));
1988
1989         return ptr;
1990 }
1991
1992 void jbd2_free(void *ptr, size_t size)
1993 {
1994         if (size == PAGE_SIZE) {
1995                 free_pages((unsigned long)ptr, 0);
1996                 return;
1997         }
1998         if (size > PAGE_SIZE) {
1999                 int order = get_order(size);
2000
2001                 if (order < 3)
2002                         free_pages((unsigned long)ptr, order);
2003                 else
2004                         vfree(ptr);
2005                 return;
2006         }
2007         kmem_cache_free(get_slab(size), ptr);
2008 };
2009
2010 /*
2011  * Journal_head storage management
2012  */
2013 static struct kmem_cache *jbd2_journal_head_cache;
2014 #ifdef CONFIG_JBD2_DEBUG
2015 static atomic_t nr_journal_heads = ATOMIC_INIT(0);
2016 #endif
2017
2018 static int journal_init_jbd2_journal_head_cache(void)
2019 {
2020         int retval;
2021
2022         J_ASSERT(jbd2_journal_head_cache == NULL);
2023         jbd2_journal_head_cache = kmem_cache_create("jbd2_journal_head",
2024                                 sizeof(struct journal_head),
2025                                 0,              /* offset */
2026                                 SLAB_TEMPORARY, /* flags */
2027                                 NULL);          /* ctor */
2028         retval = 0;
2029         if (!jbd2_journal_head_cache) {
2030                 retval = -ENOMEM;
2031                 printk(KERN_EMERG "JBD2: no memory for journal_head cache\n");
2032         }
2033         return retval;
2034 }
2035
2036 static void jbd2_journal_destroy_jbd2_journal_head_cache(void)
2037 {
2038         if (jbd2_journal_head_cache) {
2039                 kmem_cache_destroy(jbd2_journal_head_cache);
2040                 jbd2_journal_head_cache = NULL;
2041         }
2042 }
2043
2044 /*
2045  * journal_head splicing and dicing
2046  */
2047 static struct journal_head *journal_alloc_journal_head(void)
2048 {
2049         struct journal_head *ret;
2050
2051 #ifdef CONFIG_JBD2_DEBUG
2052         atomic_inc(&nr_journal_heads);
2053 #endif
2054         ret = kmem_cache_alloc(jbd2_journal_head_cache, GFP_NOFS);
2055         if (!ret) {
2056                 jbd_debug(1, "out of memory for journal_head\n");
2057                 pr_notice_ratelimited("ENOMEM in %s, retrying.\n", __func__);
2058                 while (!ret) {
2059                         yield();
2060                         ret = kmem_cache_alloc(jbd2_journal_head_cache, GFP_NOFS);
2061                 }
2062         }
2063         return ret;
2064 }
2065
2066 static void journal_free_journal_head(struct journal_head *jh)
2067 {
2068 #ifdef CONFIG_JBD2_DEBUG
2069         atomic_dec(&nr_journal_heads);
2070         memset(jh, JBD2_POISON_FREE, sizeof(*jh));
2071 #endif
2072         kmem_cache_free(jbd2_journal_head_cache, jh);
2073 }
2074
2075 /*
2076  * A journal_head is attached to a buffer_head whenever JBD has an
2077  * interest in the buffer.
2078  *
2079  * Whenever a buffer has an attached journal_head, its ->b_state:BH_JBD bit
2080  * is set.  This bit is tested in core kernel code where we need to take
2081  * JBD-specific actions.  Testing the zeroness of ->b_private is not reliable
2082  * there.
2083  *
2084  * When a buffer has its BH_JBD bit set, its ->b_count is elevated by one.
2085  *
2086  * When a buffer has its BH_JBD bit set it is immune from being released by
2087  * core kernel code, mainly via ->b_count.
2088  *
2089  * A journal_head is detached from its buffer_head when the journal_head's
2090  * b_jcount reaches zero. Running transaction (b_transaction) and checkpoint
2091  * transaction (b_cp_transaction) hold their references to b_jcount.
2092  *
2093  * Various places in the kernel want to attach a journal_head to a buffer_head
2094  * _before_ attaching the journal_head to a transaction.  To protect the
2095  * journal_head in this situation, jbd2_journal_add_journal_head elevates the
2096  * journal_head's b_jcount refcount by one.  The caller must call
2097  * jbd2_journal_put_journal_head() to undo this.
2098  *
2099  * So the typical usage would be:
2100  *
2101  *      (Attach a journal_head if needed.  Increments b_jcount)
2102  *      struct journal_head *jh = jbd2_journal_add_journal_head(bh);
2103  *      ...
2104  *      (Get another reference for transaction)
2105  *      jbd2_journal_grab_journal_head(bh);
2106  *      jh->b_transaction = xxx;
2107  *      (Put original reference)
2108  *      jbd2_journal_put_journal_head(jh);
2109  */
2110
2111 /*
2112  * Give a buffer_head a journal_head.
2113  *
2114  * May sleep.
2115  */
2116 struct journal_head *jbd2_journal_add_journal_head(struct buffer_head *bh)
2117 {
2118         struct journal_head *jh;
2119         struct journal_head *new_jh = NULL;
2120
2121 repeat:
2122         if (!buffer_jbd(bh)) {
2123                 new_jh = journal_alloc_journal_head();
2124                 memset(new_jh, 0, sizeof(*new_jh));
2125         }
2126
2127         jbd_lock_bh_journal_head(bh);
2128         if (buffer_jbd(bh)) {
2129                 jh = bh2jh(bh);
2130         } else {
2131                 J_ASSERT_BH(bh,
2132                         (atomic_read(&bh->b_count) > 0) ||
2133                         (bh->b_page && bh->b_page->mapping));
2134
2135                 if (!new_jh) {
2136                         jbd_unlock_bh_journal_head(bh);
2137                         goto repeat;
2138                 }
2139
2140                 jh = new_jh;
2141                 new_jh = NULL;          /* We consumed it */
2142                 set_buffer_jbd(bh);
2143                 bh->b_private = jh;
2144                 jh->b_bh = bh;
2145                 get_bh(bh);
2146                 BUFFER_TRACE(bh, "added journal_head");
2147         }
2148         jh->b_jcount++;
2149         jbd_unlock_bh_journal_head(bh);
2150         if (new_jh)
2151                 journal_free_journal_head(new_jh);
2152         return bh->b_private;
2153 }
2154
2155 /*
2156  * Grab a ref against this buffer_head's journal_head.  If it ended up not
2157  * having a journal_head, return NULL
2158  */
2159 struct journal_head *jbd2_journal_grab_journal_head(struct buffer_head *bh)
2160 {
2161         struct journal_head *jh = NULL;
2162
2163         jbd_lock_bh_journal_head(bh);
2164         if (buffer_jbd(bh)) {
2165                 jh = bh2jh(bh);
2166                 jh->b_jcount++;
2167         }
2168         jbd_unlock_bh_journal_head(bh);
2169         return jh;
2170 }
2171
2172 static void __journal_remove_journal_head(struct buffer_head *bh)
2173 {
2174         struct journal_head *jh = bh2jh(bh);
2175
2176         J_ASSERT_JH(jh, jh->b_jcount >= 0);
2177         J_ASSERT_JH(jh, jh->b_transaction == NULL);
2178         J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
2179         J_ASSERT_JH(jh, jh->b_cp_transaction == NULL);
2180         J_ASSERT_JH(jh, jh->b_jlist == BJ_None);
2181         J_ASSERT_BH(bh, buffer_jbd(bh));
2182         J_ASSERT_BH(bh, jh2bh(jh) == bh);
2183         BUFFER_TRACE(bh, "remove journal_head");
2184         if (jh->b_frozen_data) {
2185                 printk(KERN_WARNING "%s: freeing b_frozen_data\n", __func__);
2186                 jbd2_free(jh->b_frozen_data, bh->b_size);
2187         }
2188         if (jh->b_committed_data) {
2189                 printk(KERN_WARNING "%s: freeing b_committed_data\n", __func__);
2190                 jbd2_free(jh->b_committed_data, bh->b_size);
2191         }
2192         bh->b_private = NULL;
2193         jh->b_bh = NULL;        /* debug, really */
2194         clear_buffer_jbd(bh);
2195         journal_free_journal_head(jh);
2196 }
2197
2198 /*
2199  * Drop a reference on the passed journal_head.  If it fell to zero then
2200  * release the journal_head from the buffer_head.
2201  */
2202 void jbd2_journal_put_journal_head(struct journal_head *jh)
2203 {
2204         struct buffer_head *bh = jh2bh(jh);
2205
2206         jbd_lock_bh_journal_head(bh);
2207         J_ASSERT_JH(jh, jh->b_jcount > 0);
2208         --jh->b_jcount;
2209         if (!jh->b_jcount) {
2210                 __journal_remove_journal_head(bh);
2211                 jbd_unlock_bh_journal_head(bh);
2212                 __brelse(bh);
2213         } else
2214                 jbd_unlock_bh_journal_head(bh);
2215 }
2216
2217 /*
2218  * Initialize jbd inode head
2219  */
2220 void jbd2_journal_init_jbd_inode(struct jbd2_inode *jinode, struct inode *inode)
2221 {
2222         jinode->i_transaction = NULL;
2223         jinode->i_next_transaction = NULL;
2224         jinode->i_vfs_inode = inode;
2225         jinode->i_flags = 0;
2226         INIT_LIST_HEAD(&jinode->i_list);
2227 }
2228
2229 /*
2230  * Function to be called before we start removing inode from memory (i.e.,
2231  * clear_inode() is a fine place to be called from). It removes inode from
2232  * transaction's lists.
2233  */
2234 void jbd2_journal_release_jbd_inode(journal_t *journal,
2235                                     struct jbd2_inode *jinode)
2236 {
2237         if (!journal)
2238                 return;
2239 restart:
2240         spin_lock(&journal->j_list_lock);
2241         /* Is commit writing out inode - we have to wait */
2242         if (test_bit(__JI_COMMIT_RUNNING, &jinode->i_flags)) {
2243                 wait_queue_head_t *wq;
2244                 DEFINE_WAIT_BIT(wait, &jinode->i_flags, __JI_COMMIT_RUNNING);
2245                 wq = bit_waitqueue(&jinode->i_flags, __JI_COMMIT_RUNNING);
2246                 prepare_to_wait(wq, &wait.wait, TASK_UNINTERRUPTIBLE);
2247                 spin_unlock(&journal->j_list_lock);
2248                 schedule();
2249                 finish_wait(wq, &wait.wait);
2250                 goto restart;
2251         }
2252
2253         if (jinode->i_transaction) {
2254                 list_del(&jinode->i_list);
2255                 jinode->i_transaction = NULL;
2256         }
2257         spin_unlock(&journal->j_list_lock);
2258 }
2259
2260 /*
2261  * debugfs tunables
2262  */
2263 #ifdef CONFIG_JBD2_DEBUG
2264 u8 jbd2_journal_enable_debug __read_mostly;
2265 EXPORT_SYMBOL(jbd2_journal_enable_debug);
2266
2267 #define JBD2_DEBUG_NAME "jbd2-debug"
2268
2269 static struct dentry *jbd2_debugfs_dir;
2270 static struct dentry *jbd2_debug;
2271
2272 static void __init jbd2_create_debugfs_entry(void)
2273 {
2274         jbd2_debugfs_dir = debugfs_create_dir("jbd2", NULL);
2275         if (jbd2_debugfs_dir)
2276                 jbd2_debug = debugfs_create_u8(JBD2_DEBUG_NAME,
2277                                                S_IRUGO | S_IWUSR,
2278                                                jbd2_debugfs_dir,
2279                                                &jbd2_journal_enable_debug);
2280 }
2281
2282 static void __exit jbd2_remove_debugfs_entry(void)
2283 {
2284         debugfs_remove(jbd2_debug);
2285         debugfs_remove(jbd2_debugfs_dir);
2286 }
2287
2288 #else
2289
2290 static void __init jbd2_create_debugfs_entry(void)
2291 {
2292 }
2293
2294 static void __exit jbd2_remove_debugfs_entry(void)
2295 {
2296 }
2297
2298 #endif
2299
2300 #ifdef CONFIG_PROC_FS
2301
2302 #define JBD2_STATS_PROC_NAME "fs/jbd2"
2303
2304 static void __init jbd2_create_jbd_stats_proc_entry(void)
2305 {
2306         proc_jbd2_stats = proc_mkdir(JBD2_STATS_PROC_NAME, NULL);
2307 }
2308
2309 static void __exit jbd2_remove_jbd_stats_proc_entry(void)
2310 {
2311         if (proc_jbd2_stats)
2312                 remove_proc_entry(JBD2_STATS_PROC_NAME, NULL);
2313 }
2314
2315 #else
2316
2317 #define jbd2_create_jbd_stats_proc_entry() do {} while (0)
2318 #define jbd2_remove_jbd_stats_proc_entry() do {} while (0)
2319
2320 #endif
2321
2322 struct kmem_cache *jbd2_handle_cache, *jbd2_inode_cache;
2323
2324 static int __init journal_init_handle_cache(void)
2325 {
2326         jbd2_handle_cache = KMEM_CACHE(jbd2_journal_handle, SLAB_TEMPORARY);
2327         if (jbd2_handle_cache == NULL) {
2328                 printk(KERN_EMERG "JBD2: failed to create handle cache\n");
2329                 return -ENOMEM;
2330         }
2331         jbd2_inode_cache = KMEM_CACHE(jbd2_inode, 0);
2332         if (jbd2_inode_cache == NULL) {
2333                 printk(KERN_EMERG "JBD2: failed to create inode cache\n");
2334                 kmem_cache_destroy(jbd2_handle_cache);
2335                 return -ENOMEM;
2336         }
2337         return 0;
2338 }
2339
2340 static void jbd2_journal_destroy_handle_cache(void)
2341 {
2342         if (jbd2_handle_cache)
2343                 kmem_cache_destroy(jbd2_handle_cache);
2344         if (jbd2_inode_cache)
2345                 kmem_cache_destroy(jbd2_inode_cache);
2346
2347 }
2348
2349 /*
2350  * Module startup and shutdown
2351  */
2352
2353 static int __init journal_init_caches(void)
2354 {
2355         int ret;
2356
2357         ret = jbd2_journal_init_revoke_caches();
2358         if (ret == 0)
2359                 ret = journal_init_jbd2_journal_head_cache();
2360         if (ret == 0)
2361                 ret = journal_init_handle_cache();
2362         return ret;
2363 }
2364
2365 static void jbd2_journal_destroy_caches(void)
2366 {
2367         jbd2_journal_destroy_revoke_caches();
2368         jbd2_journal_destroy_jbd2_journal_head_cache();
2369         jbd2_journal_destroy_handle_cache();
2370         jbd2_journal_destroy_slabs();
2371 }
2372
2373 static int __init journal_init(void)
2374 {
2375         int ret;
2376
2377         BUILD_BUG_ON(sizeof(struct journal_superblock_s) != 1024);
2378
2379         ret = journal_init_caches();
2380         if (ret == 0) {
2381                 jbd2_create_debugfs_entry();
2382                 jbd2_create_jbd_stats_proc_entry();
2383         } else {
2384                 jbd2_journal_destroy_caches();
2385         }
2386         return ret;
2387 }
2388
2389 static void __exit journal_exit(void)
2390 {
2391 #ifdef CONFIG_JBD2_DEBUG
2392         int n = atomic_read(&nr_journal_heads);
2393         if (n)
2394                 printk(KERN_EMERG "JBD2: leaked %d journal_heads!\n", n);
2395 #endif
2396         jbd2_remove_debugfs_entry();
2397         jbd2_remove_jbd_stats_proc_entry();
2398         jbd2_journal_destroy_caches();
2399 }
2400
2401 MODULE_LICENSE("GPL");
2402 module_init(journal_init);
2403 module_exit(journal_exit);
2404